4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2022 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
26 * Copyright (c) 2019, Klara Inc.
27 * Copyright (c) 2019, Allan Jude
28 * Copyright (c) 2021, Datto, Inc.
31 #include <sys/sysmacros.h>
32 #include <sys/zfs_context.h>
33 #include <sys/fm/fs/zfs.h>
36 #include <sys/spa_impl.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/vdev_trim.h>
39 #include <sys/zio_impl.h>
40 #include <sys/zio_compress.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/dmu_objset.h>
46 #include <sys/blkptr.h>
47 #include <sys/zfeature.h>
48 #include <sys/dsl_scan.h>
49 #include <sys/metaslab_impl.h>
51 #include <sys/trace_zfs.h>
53 #include <sys/dsl_crypt.h>
57 * ==========================================================================
58 * I/O type descriptions
59 * ==========================================================================
61 const char *const zio_type_name
[ZIO_TYPES
] = {
63 * Note: Linux kernel thread name length is limited
64 * so these names will differ from upstream open zfs.
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
69 int zio_dva_throttle_enabled
= B_TRUE
;
70 static int zio_deadman_log_all
= B_FALSE
;
73 * ==========================================================================
75 * ==========================================================================
77 static kmem_cache_t
*zio_cache
;
78 static kmem_cache_t
*zio_link_cache
;
79 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
80 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
81 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
82 static uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
83 static uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
86 /* Mark IOs as "slow" if they take longer than 30 seconds */
87 static uint_t zio_slow_io_ms
= (30 * MILLISEC
);
89 #define BP_SPANB(indblkshift, level) \
90 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
91 #define COMPARE_META_LEVEL 0x80000000ul
93 * The following actions directly effect the spa's sync-to-convergence logic.
94 * The values below define the sync pass when we start performing the action.
95 * Care should be taken when changing these values as they directly impact
96 * spa_sync() performance. Tuning these values may introduce subtle performance
97 * pathologies and should only be done in the context of performance analysis.
98 * These tunables will eventually be removed and replaced with #defines once
99 * enough analysis has been done to determine optimal values.
101 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
102 * regular blocks are not deferred.
104 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
105 * compression (including of metadata). In practice, we don't have this
106 * many sync passes, so this has no effect.
108 * The original intent was that disabling compression would help the sync
109 * passes to converge. However, in practice disabling compression increases
110 * the average number of sync passes, because when we turn compression off, a
111 * lot of block's size will change and thus we have to re-allocate (not
112 * overwrite) them. It also increases the number of 128KB allocations (e.g.
113 * for indirect blocks and spacemaps) because these will not be compressed.
114 * The 128K allocations are especially detrimental to performance on highly
115 * fragmented systems, which may have very few free segments of this size,
116 * and may need to load new metaslabs to satisfy 128K allocations.
119 /* defer frees starting in this pass */
120 uint_t zfs_sync_pass_deferred_free
= 2;
122 /* don't compress starting in this pass */
123 static uint_t zfs_sync_pass_dont_compress
= 8;
125 /* rewrite new bps starting in this pass */
126 static uint_t zfs_sync_pass_rewrite
= 2;
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 * Enable smaller cores by excluding metadata
136 * allocations as well.
138 int zio_exclude_metadata
= 0;
139 static int zio_requeue_io_start_cut_in_line
= 1;
142 static const int zio_buf_debug_limit
= 16384;
144 static const int zio_buf_debug_limit
= 0;
147 static inline void __zio_execute(zio_t
*zio
);
149 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
156 zio_cache
= kmem_cache_create("zio_cache",
157 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
158 zio_link_cache
= kmem_cache_create("zio_link_cache",
159 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
162 * For small buffers, we want a cache for each multiple of
163 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
164 * for each quarter-power of 2.
166 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
167 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
170 size_t data_cflags
, cflags
;
172 data_cflags
= KMC_NODEBUG
;
173 cflags
= (zio_exclude_metadata
|| size
> zio_buf_debug_limit
) ?
181 * If we are using watchpoints, put each buffer on its own page,
182 * to eliminate the performance overhead of trapping to the
183 * kernel when modifying a non-watched buffer that shares the
184 * page with a watched buffer.
186 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
189 * Here's the problem - on 4K native devices in userland on
190 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
191 * will fail with EINVAL, causing zdb (and others) to coredump.
192 * Since userland probably doesn't need optimized buffer caches,
193 * we just force 4K alignment on everything.
195 align
= 8 * SPA_MINBLOCKSIZE
;
197 if (size
< PAGESIZE
) {
198 align
= SPA_MINBLOCKSIZE
;
199 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
206 if (cflags
== data_cflags
) {
208 * Resulting kmem caches would be identical.
209 * Save memory by creating only one.
211 (void) snprintf(name
, sizeof (name
),
212 "zio_buf_comb_%lu", (ulong_t
)size
);
213 zio_buf_cache
[c
] = kmem_cache_create(name
,
214 size
, align
, NULL
, NULL
, NULL
, NULL
, NULL
,
216 zio_data_buf_cache
[c
] = zio_buf_cache
[c
];
219 (void) snprintf(name
, sizeof (name
), "zio_buf_%lu",
221 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
222 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
224 (void) snprintf(name
, sizeof (name
), "zio_data_buf_%lu",
226 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
227 align
, NULL
, NULL
, NULL
, NULL
, NULL
, data_cflags
);
232 ASSERT(zio_buf_cache
[c
] != NULL
);
233 if (zio_buf_cache
[c
- 1] == NULL
)
234 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
236 ASSERT(zio_data_buf_cache
[c
] != NULL
);
237 if (zio_data_buf_cache
[c
- 1] == NULL
)
238 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
249 size_t n
= SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
;
251 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
252 for (size_t i
= 0; i
< n
; i
++) {
253 if (zio_buf_cache_allocs
[i
] != zio_buf_cache_frees
[i
])
254 (void) printf("zio_fini: [%d] %llu != %llu\n",
255 (int)((i
+ 1) << SPA_MINBLOCKSHIFT
),
256 (long long unsigned)zio_buf_cache_allocs
[i
],
257 (long long unsigned)zio_buf_cache_frees
[i
]);
262 * The same kmem cache can show up multiple times in both zio_buf_cache
263 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
266 for (size_t i
= 0; i
< n
; i
++) {
267 kmem_cache_t
*cache
= zio_buf_cache
[i
];
270 for (size_t j
= i
; j
< n
; j
++) {
271 if (cache
== zio_buf_cache
[j
])
272 zio_buf_cache
[j
] = NULL
;
273 if (cache
== zio_data_buf_cache
[j
])
274 zio_data_buf_cache
[j
] = NULL
;
276 kmem_cache_destroy(cache
);
279 for (size_t i
= 0; i
< n
; i
++) {
280 kmem_cache_t
*cache
= zio_data_buf_cache
[i
];
283 for (size_t j
= i
; j
< n
; j
++) {
284 if (cache
== zio_data_buf_cache
[j
])
285 zio_data_buf_cache
[j
] = NULL
;
287 kmem_cache_destroy(cache
);
290 for (size_t i
= 0; i
< n
; i
++) {
291 VERIFY3P(zio_buf_cache
[i
], ==, NULL
);
292 VERIFY3P(zio_data_buf_cache
[i
], ==, NULL
);
295 kmem_cache_destroy(zio_link_cache
);
296 kmem_cache_destroy(zio_cache
);
304 * ==========================================================================
305 * Allocate and free I/O buffers
306 * ==========================================================================
310 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
311 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
312 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
313 * excess / transient data in-core during a crashdump.
316 zio_buf_alloc(size_t size
)
318 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
320 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
321 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
322 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
325 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
329 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
330 * crashdump if the kernel panics. This exists so that we will limit the amount
331 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
332 * of kernel heap dumped to disk when the kernel panics)
335 zio_data_buf_alloc(size_t size
)
337 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
339 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
341 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
345 zio_buf_free(void *buf
, size_t size
)
347 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
349 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
350 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
351 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
354 kmem_cache_free(zio_buf_cache
[c
], buf
);
358 zio_data_buf_free(void *buf
, size_t size
)
360 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
362 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
364 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
368 zio_abd_free(void *abd
, size_t size
)
371 abd_free((abd_t
*)abd
);
375 * ==========================================================================
376 * Push and pop I/O transform buffers
377 * ==========================================================================
380 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
381 zio_transform_func_t
*transform
)
383 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
385 zt
->zt_orig_abd
= zio
->io_abd
;
386 zt
->zt_orig_size
= zio
->io_size
;
387 zt
->zt_bufsize
= bufsize
;
388 zt
->zt_transform
= transform
;
390 zt
->zt_next
= zio
->io_transform_stack
;
391 zio
->io_transform_stack
= zt
;
398 zio_pop_transforms(zio_t
*zio
)
402 while ((zt
= zio
->io_transform_stack
) != NULL
) {
403 if (zt
->zt_transform
!= NULL
)
404 zt
->zt_transform(zio
,
405 zt
->zt_orig_abd
, zt
->zt_orig_size
);
407 if (zt
->zt_bufsize
!= 0)
408 abd_free(zio
->io_abd
);
410 zio
->io_abd
= zt
->zt_orig_abd
;
411 zio
->io_size
= zt
->zt_orig_size
;
412 zio
->io_transform_stack
= zt
->zt_next
;
414 kmem_free(zt
, sizeof (zio_transform_t
));
419 * ==========================================================================
420 * I/O transform callbacks for subblocks, decompression, and decryption
421 * ==========================================================================
424 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
426 ASSERT(zio
->io_size
> size
);
428 if (zio
->io_type
== ZIO_TYPE_READ
)
429 abd_copy(data
, zio
->io_abd
, size
);
433 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
435 if (zio
->io_error
== 0) {
436 void *tmp
= abd_borrow_buf(data
, size
);
437 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
438 zio
->io_abd
, tmp
, zio
->io_size
, size
,
439 &zio
->io_prop
.zp_complevel
);
440 abd_return_buf_copy(data
, tmp
, size
);
442 if (zio_injection_enabled
&& ret
== 0)
443 ret
= zio_handle_fault_injection(zio
, EINVAL
);
446 zio
->io_error
= SET_ERROR(EIO
);
451 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
455 blkptr_t
*bp
= zio
->io_bp
;
456 spa_t
*spa
= zio
->io_spa
;
457 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
458 uint64_t lsize
= BP_GET_LSIZE(bp
);
459 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
460 uint8_t salt
[ZIO_DATA_SALT_LEN
];
461 uint8_t iv
[ZIO_DATA_IV_LEN
];
462 uint8_t mac
[ZIO_DATA_MAC_LEN
];
463 boolean_t no_crypt
= B_FALSE
;
465 ASSERT(BP_USES_CRYPT(bp
));
466 ASSERT3U(size
, !=, 0);
468 if (zio
->io_error
!= 0)
472 * Verify the cksum of MACs stored in an indirect bp. It will always
473 * be possible to verify this since it does not require an encryption
476 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
477 zio_crypt_decode_mac_bp(bp
, mac
);
479 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
481 * We haven't decompressed the data yet, but
482 * zio_crypt_do_indirect_mac_checksum() requires
483 * decompressed data to be able to parse out the MACs
484 * from the indirect block. We decompress it now and
485 * throw away the result after we are finished.
487 tmp
= zio_buf_alloc(lsize
);
488 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
489 zio
->io_abd
, tmp
, zio
->io_size
, lsize
,
490 &zio
->io_prop
.zp_complevel
);
492 ret
= SET_ERROR(EIO
);
495 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
496 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
497 zio_buf_free(tmp
, lsize
);
499 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
500 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
502 abd_copy(data
, zio
->io_abd
, size
);
504 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
505 ret
= zio_handle_decrypt_injection(spa
,
506 &zio
->io_bookmark
, ot
, ECKSUM
);
515 * If this is an authenticated block, just check the MAC. It would be
516 * nice to separate this out into its own flag, but when this was done,
517 * we had run out of bits in what is now zio_flag_t. Future cleanup
518 * could make this a flag bit.
520 if (BP_IS_AUTHENTICATED(bp
)) {
521 if (ot
== DMU_OT_OBJSET
) {
522 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
523 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
525 zio_crypt_decode_mac_bp(bp
, mac
);
526 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
527 zio
->io_abd
, size
, mac
);
528 if (zio_injection_enabled
&& ret
== 0) {
529 ret
= zio_handle_decrypt_injection(spa
,
530 &zio
->io_bookmark
, ot
, ECKSUM
);
533 abd_copy(data
, zio
->io_abd
, size
);
541 zio_crypt_decode_params_bp(bp
, salt
, iv
);
543 if (ot
== DMU_OT_INTENT_LOG
) {
544 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
545 zio_crypt_decode_mac_zil(tmp
, mac
);
546 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
548 zio_crypt_decode_mac_bp(bp
, mac
);
551 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
552 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
553 zio
->io_abd
, &no_crypt
);
555 abd_copy(data
, zio
->io_abd
, size
);
563 /* assert that the key was found unless this was speculative */
564 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
567 * If there was a decryption / authentication error return EIO as
568 * the io_error. If this was not a speculative zio, create an ereport.
571 zio
->io_error
= SET_ERROR(EIO
);
572 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
573 spa_log_error(spa
, &zio
->io_bookmark
,
574 &zio
->io_bp
->blk_birth
);
575 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
576 spa
, NULL
, &zio
->io_bookmark
, zio
, 0);
584 * ==========================================================================
585 * I/O parent/child relationships and pipeline interlocks
586 * ==========================================================================
589 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
591 list_t
*pl
= &cio
->io_parent_list
;
593 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
597 ASSERT((*zl
)->zl_child
== cio
);
598 return ((*zl
)->zl_parent
);
602 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
604 list_t
*cl
= &pio
->io_child_list
;
606 ASSERT(MUTEX_HELD(&pio
->io_lock
));
608 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
612 ASSERT((*zl
)->zl_parent
== pio
);
613 return ((*zl
)->zl_child
);
617 zio_unique_parent(zio_t
*cio
)
619 zio_link_t
*zl
= NULL
;
620 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
622 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
627 zio_add_child(zio_t
*pio
, zio_t
*cio
)
629 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
632 * Logical I/Os can have logical, gang, or vdev children.
633 * Gang I/Os can have gang or vdev children.
634 * Vdev I/Os can only have vdev children.
635 * The following ASSERT captures all of these constraints.
637 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
642 mutex_enter(&pio
->io_lock
);
643 mutex_enter(&cio
->io_lock
);
645 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
647 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
648 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
650 list_insert_head(&pio
->io_child_list
, zl
);
651 list_insert_head(&cio
->io_parent_list
, zl
);
653 mutex_exit(&cio
->io_lock
);
654 mutex_exit(&pio
->io_lock
);
658 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
660 ASSERT(zl
->zl_parent
== pio
);
661 ASSERT(zl
->zl_child
== cio
);
663 mutex_enter(&pio
->io_lock
);
664 mutex_enter(&cio
->io_lock
);
666 list_remove(&pio
->io_child_list
, zl
);
667 list_remove(&cio
->io_parent_list
, zl
);
669 mutex_exit(&cio
->io_lock
);
670 mutex_exit(&pio
->io_lock
);
671 kmem_cache_free(zio_link_cache
, zl
);
675 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
677 boolean_t waiting
= B_FALSE
;
679 mutex_enter(&zio
->io_lock
);
680 ASSERT(zio
->io_stall
== NULL
);
681 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
682 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
685 uint64_t *countp
= &zio
->io_children
[c
][wait
];
688 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
689 zio
->io_stall
= countp
;
694 mutex_exit(&zio
->io_lock
);
698 __attribute__((always_inline
))
700 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
701 zio_t
**next_to_executep
)
703 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
704 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
706 mutex_enter(&pio
->io_lock
);
707 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
708 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
709 pio
->io_reexecute
|= zio
->io_reexecute
;
710 ASSERT3U(*countp
, >, 0);
714 if (*countp
== 0 && pio
->io_stall
== countp
) {
715 zio_taskq_type_t type
=
716 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
718 pio
->io_stall
= NULL
;
719 mutex_exit(&pio
->io_lock
);
722 * If we can tell the caller to execute this parent next, do
723 * so. We only do this if the parent's zio type matches the
724 * child's type. Otherwise dispatch the parent zio in its
727 * Having the caller execute the parent when possible reduces
728 * locking on the zio taskq's, reduces context switch
729 * overhead, and has no recursion penalty. Note that one
730 * read from disk typically causes at least 3 zio's: a
731 * zio_null(), the logical zio_read(), and then a physical
732 * zio. When the physical ZIO completes, we are able to call
733 * zio_done() on all 3 of these zio's from one invocation of
734 * zio_execute() by returning the parent back to
735 * zio_execute(). Since the parent isn't executed until this
736 * thread returns back to zio_execute(), the caller should do
739 * In other cases, dispatching the parent prevents
740 * overflowing the stack when we have deeply nested
741 * parent-child relationships, as we do with the "mega zio"
742 * of writes for spa_sync(), and the chain of ZIL blocks.
744 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
&&
745 pio
->io_type
== zio
->io_type
) {
746 *next_to_executep
= pio
;
748 zio_taskq_dispatch(pio
, type
, B_FALSE
);
751 mutex_exit(&pio
->io_lock
);
756 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
758 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
759 zio
->io_error
= zio
->io_child_error
[c
];
763 zio_bookmark_compare(const void *x1
, const void *x2
)
765 const zio_t
*z1
= x1
;
766 const zio_t
*z2
= x2
;
768 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
770 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
773 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
775 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
778 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
780 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
783 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
785 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
797 * ==========================================================================
798 * Create the various types of I/O (read, write, free, etc)
799 * ==========================================================================
802 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
803 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
804 void *private, zio_type_t type
, zio_priority_t priority
,
805 zio_flag_t flags
, vdev_t
*vd
, uint64_t offset
,
806 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
807 enum zio_stage pipeline
)
811 IMPLY(type
!= ZIO_TYPE_TRIM
, psize
<= SPA_MAXBLOCKSIZE
);
812 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
813 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
815 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
816 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
817 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
819 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
821 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
822 memset(zio
, 0, sizeof (zio_t
));
824 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
825 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
827 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
828 offsetof(zio_link_t
, zl_parent_node
));
829 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
830 offsetof(zio_link_t
, zl_child_node
));
831 metaslab_trace_init(&zio
->io_alloc_list
);
834 zio
->io_child_type
= ZIO_CHILD_VDEV
;
835 else if (flags
& ZIO_FLAG_GANG_CHILD
)
836 zio
->io_child_type
= ZIO_CHILD_GANG
;
837 else if (flags
& ZIO_FLAG_DDT_CHILD
)
838 zio
->io_child_type
= ZIO_CHILD_DDT
;
840 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
843 zio
->io_bp
= (blkptr_t
*)bp
;
844 zio
->io_bp_copy
= *bp
;
845 zio
->io_bp_orig
= *bp
;
846 if (type
!= ZIO_TYPE_WRITE
||
847 zio
->io_child_type
== ZIO_CHILD_DDT
)
848 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
849 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
850 zio
->io_logical
= zio
;
851 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
852 pipeline
|= ZIO_GANG_STAGES
;
858 zio
->io_private
= private;
860 zio
->io_priority
= priority
;
862 zio
->io_offset
= offset
;
863 zio
->io_orig_abd
= zio
->io_abd
= data
;
864 zio
->io_orig_size
= zio
->io_size
= psize
;
865 zio
->io_lsize
= lsize
;
866 zio
->io_orig_flags
= zio
->io_flags
= flags
;
867 zio
->io_orig_stage
= zio
->io_stage
= stage
;
868 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
869 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
871 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
872 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
875 zio
->io_bookmark
= *zb
;
878 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
879 if (zio
->io_logical
== NULL
)
880 zio
->io_logical
= pio
->io_logical
;
881 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
882 zio
->io_gang_leader
= pio
->io_gang_leader
;
883 zio_add_child(pio
, zio
);
886 taskq_init_ent(&zio
->io_tqent
);
892 zio_destroy(zio_t
*zio
)
894 metaslab_trace_fini(&zio
->io_alloc_list
);
895 list_destroy(&zio
->io_parent_list
);
896 list_destroy(&zio
->io_child_list
);
897 mutex_destroy(&zio
->io_lock
);
898 cv_destroy(&zio
->io_cv
);
899 kmem_cache_free(zio_cache
, zio
);
903 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
904 void *private, zio_flag_t flags
)
908 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
909 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
910 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
916 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, zio_flag_t flags
)
918 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
922 zfs_blkptr_verify_log(spa_t
*spa
, const blkptr_t
*bp
,
923 enum blk_verify_flag blk_verify
, const char *fmt
, ...)
929 (void) vsnprintf(buf
, sizeof (buf
), fmt
, adx
);
932 zfs_dbgmsg("bad blkptr at %px: "
933 "DVA[0]=%#llx/%#llx "
934 "DVA[1]=%#llx/%#llx "
935 "DVA[2]=%#llx/%#llx "
941 "cksum=%#llx/%#llx/%#llx/%#llx",
943 (long long)bp
->blk_dva
[0].dva_word
[0],
944 (long long)bp
->blk_dva
[0].dva_word
[1],
945 (long long)bp
->blk_dva
[1].dva_word
[0],
946 (long long)bp
->blk_dva
[1].dva_word
[1],
947 (long long)bp
->blk_dva
[2].dva_word
[0],
948 (long long)bp
->blk_dva
[2].dva_word
[1],
949 (long long)bp
->blk_prop
,
950 (long long)bp
->blk_pad
[0],
951 (long long)bp
->blk_pad
[1],
952 (long long)bp
->blk_phys_birth
,
953 (long long)bp
->blk_birth
,
954 (long long)bp
->blk_fill
,
955 (long long)bp
->blk_cksum
.zc_word
[0],
956 (long long)bp
->blk_cksum
.zc_word
[1],
957 (long long)bp
->blk_cksum
.zc_word
[2],
958 (long long)bp
->blk_cksum
.zc_word
[3]);
959 switch (blk_verify
) {
960 case BLK_VERIFY_HALT
:
961 zfs_panic_recover("%s: %s", spa_name(spa
), buf
);
964 zfs_dbgmsg("%s: %s", spa_name(spa
), buf
);
966 case BLK_VERIFY_ONLY
:
974 * Verify the block pointer fields contain reasonable values. This means
975 * it only contains known object types, checksum/compression identifiers,
976 * block sizes within the maximum allowed limits, valid DVAs, etc.
978 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
979 * argument controls the behavior when an invalid field is detected.
981 * Values for blk_verify_flag:
982 * BLK_VERIFY_ONLY: evaluate the block
983 * BLK_VERIFY_LOG: evaluate the block and log problems
984 * BLK_VERIFY_HALT: call zfs_panic_recover on error
986 * Values for blk_config_flag:
987 * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
988 * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
989 * obtained for reader
990 * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
994 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
,
995 enum blk_config_flag blk_config
, enum blk_verify_flag blk_verify
)
999 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
1000 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1001 "blkptr at %px has invalid TYPE %llu",
1002 bp
, (longlong_t
)BP_GET_TYPE(bp
));
1004 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
) {
1005 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1006 "blkptr at %px has invalid CHECKSUM %llu",
1007 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
1009 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
) {
1010 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1011 "blkptr at %px has invalid COMPRESS %llu",
1012 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
1014 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
1015 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1016 "blkptr at %px has invalid LSIZE %llu",
1017 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
1019 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
1020 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1021 "blkptr at %px has invalid PSIZE %llu",
1022 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
1025 if (BP_IS_EMBEDDED(bp
)) {
1026 if (BPE_GET_ETYPE(bp
) >= NUM_BP_EMBEDDED_TYPES
) {
1027 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1028 "blkptr at %px has invalid ETYPE %llu",
1029 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
1034 * Do not verify individual DVAs if the config is not trusted. This
1035 * will be done once the zio is executed in vdev_mirror_map_alloc.
1037 if (!spa
->spa_trust_config
)
1038 return (errors
== 0);
1040 switch (blk_config
) {
1041 case BLK_CONFIG_HELD
:
1042 ASSERT(spa_config_held(spa
, SCL_VDEV
, RW_WRITER
));
1044 case BLK_CONFIG_NEEDED
:
1045 spa_config_enter(spa
, SCL_VDEV
, bp
, RW_READER
);
1047 case BLK_CONFIG_SKIP
:
1048 return (errors
== 0);
1050 panic("invalid blk_config %u", blk_config
);
1054 * Pool-specific checks.
1056 * Note: it would be nice to verify that the blk_birth and
1057 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
1058 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1059 * that are in the log) to be arbitrarily large.
1061 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
1062 const dva_t
*dva
= &bp
->blk_dva
[i
];
1063 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1065 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
1066 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1067 "blkptr at %px DVA %u has invalid VDEV %llu",
1068 bp
, i
, (longlong_t
)vdevid
);
1071 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1073 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1074 "blkptr at %px DVA %u has invalid VDEV %llu",
1075 bp
, i
, (longlong_t
)vdevid
);
1078 if (vd
->vdev_ops
== &vdev_hole_ops
) {
1079 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1080 "blkptr at %px DVA %u has hole VDEV %llu",
1081 bp
, i
, (longlong_t
)vdevid
);
1084 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1086 * "missing" vdevs are valid during import, but we
1087 * don't have their detailed info (e.g. asize), so
1088 * we can't perform any more checks on them.
1092 uint64_t offset
= DVA_GET_OFFSET(dva
);
1093 uint64_t asize
= DVA_GET_ASIZE(dva
);
1094 if (DVA_GET_GANG(dva
))
1095 asize
= vdev_gang_header_asize(vd
);
1096 if (offset
+ asize
> vd
->vdev_asize
) {
1097 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1098 "blkptr at %px DVA %u has invalid OFFSET %llu",
1099 bp
, i
, (longlong_t
)offset
);
1102 if (blk_config
== BLK_CONFIG_NEEDED
)
1103 spa_config_exit(spa
, SCL_VDEV
, bp
);
1105 return (errors
== 0);
1109 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
1112 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1114 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
1117 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1121 if (vd
->vdev_ops
== &vdev_hole_ops
)
1124 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1128 uint64_t offset
= DVA_GET_OFFSET(dva
);
1129 uint64_t asize
= DVA_GET_ASIZE(dva
);
1131 if (DVA_GET_GANG(dva
))
1132 asize
= vdev_gang_header_asize(vd
);
1133 if (offset
+ asize
> vd
->vdev_asize
)
1140 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
1141 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
1142 zio_priority_t priority
, zio_flag_t flags
, const zbookmark_phys_t
*zb
)
1146 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
1147 data
, size
, size
, done
, private,
1148 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1149 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1150 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1156 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1157 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1158 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1159 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1160 zio_flag_t flags
, const zbookmark_phys_t
*zb
)
1164 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1165 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1166 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1167 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1168 DMU_OT_IS_VALID(zp
->zp_type
) &&
1169 zp
->zp_level
< 32 &&
1170 zp
->zp_copies
> 0 &&
1171 zp
->zp_copies
<= spa_max_replication(spa
));
1173 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1174 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1175 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1176 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1178 zio
->io_ready
= ready
;
1179 zio
->io_children_ready
= children_ready
;
1183 * Data can be NULL if we are going to call zio_write_override() to
1184 * provide the already-allocated BP. But we may need the data to
1185 * verify a dedup hit (if requested). In this case, don't try to
1186 * dedup (just take the already-allocated BP verbatim). Encrypted
1187 * dedup blocks need data as well so we also disable dedup in this
1191 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1192 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1199 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1200 uint64_t size
, zio_done_func_t
*done
, void *private,
1201 zio_priority_t priority
, zio_flag_t flags
, zbookmark_phys_t
*zb
)
1205 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1206 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1207 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1213 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
,
1216 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1217 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1218 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1219 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1220 ASSERT(!brtwrite
|| !nopwrite
);
1223 * We must reset the io_prop to match the values that existed
1224 * when the bp was first written by dmu_sync() keeping in mind
1225 * that nopwrite and dedup are mutually exclusive.
1227 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1228 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1229 zio
->io_prop
.zp_brtwrite
= brtwrite
;
1230 zio
->io_prop
.zp_copies
= copies
;
1231 zio
->io_bp_override
= bp
;
1235 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1238 (void) zfs_blkptr_verify(spa
, bp
, BLK_CONFIG_NEEDED
, BLK_VERIFY_HALT
);
1241 * The check for EMBEDDED is a performance optimization. We
1242 * process the free here (by ignoring it) rather than
1243 * putting it on the list and then processing it in zio_free_sync().
1245 if (BP_IS_EMBEDDED(bp
))
1249 * Frees that are for the currently-syncing txg, are not going to be
1250 * deferred, and which will not need to do a read (i.e. not GANG or
1251 * DEDUP), can be processed immediately. Otherwise, put them on the
1252 * in-memory list for later processing.
1254 * Note that we only defer frees after zfs_sync_pass_deferred_free
1255 * when the log space map feature is disabled. [see relevant comment
1256 * in spa_sync_iterate_to_convergence()]
1258 if (BP_IS_GANG(bp
) ||
1260 txg
!= spa
->spa_syncing_txg
||
1261 (spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
&&
1262 !spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
)) ||
1263 brt_maybe_exists(spa
, bp
)) {
1264 metaslab_check_free(spa
, bp
);
1265 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1267 VERIFY3P(zio_free_sync(NULL
, spa
, txg
, bp
, 0), ==, NULL
);
1272 * To improve performance, this function may return NULL if we were able
1273 * to do the free immediately. This avoids the cost of creating a zio
1274 * (and linking it to the parent, etc).
1277 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1280 ASSERT(!BP_IS_HOLE(bp
));
1281 ASSERT(spa_syncing_txg(spa
) == txg
);
1283 if (BP_IS_EMBEDDED(bp
))
1286 metaslab_check_free(spa
, bp
);
1288 dsl_scan_freed(spa
, bp
);
1290 if (BP_IS_GANG(bp
) ||
1292 brt_maybe_exists(spa
, bp
)) {
1294 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1295 * block header, the DDT or the BRT), so issue them
1296 * asynchronously so that this thread is not tied up.
1298 enum zio_stage stage
=
1299 ZIO_FREE_PIPELINE
| ZIO_STAGE_ISSUE_ASYNC
;
1301 return (zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1302 BP_GET_PSIZE(bp
), NULL
, NULL
,
1303 ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1304 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
));
1306 metaslab_free(spa
, bp
, txg
, B_FALSE
);
1312 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1313 zio_done_func_t
*done
, void *private, zio_flag_t flags
)
1317 (void) zfs_blkptr_verify(spa
, bp
, (flags
& ZIO_FLAG_CONFIG_WRITER
) ?
1318 BLK_CONFIG_HELD
: BLK_CONFIG_NEEDED
, BLK_VERIFY_HALT
);
1320 if (BP_IS_EMBEDDED(bp
))
1321 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1324 * A claim is an allocation of a specific block. Claims are needed
1325 * to support immediate writes in the intent log. The issue is that
1326 * immediate writes contain committed data, but in a txg that was
1327 * *not* committed. Upon opening the pool after an unclean shutdown,
1328 * the intent log claims all blocks that contain immediate write data
1329 * so that the SPA knows they're in use.
1331 * All claims *must* be resolved in the first txg -- before the SPA
1332 * starts allocating blocks -- so that nothing is allocated twice.
1333 * If txg == 0 we just verify that the block is claimable.
1335 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
1336 spa_min_claim_txg(spa
));
1337 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1338 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(8) */
1340 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1341 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1342 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1343 ASSERT0(zio
->io_queued_timestamp
);
1349 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1350 zio_done_func_t
*done
, void *private, zio_flag_t flags
)
1355 if (vd
->vdev_children
== 0) {
1356 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1357 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1358 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1362 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1364 for (c
= 0; c
< vd
->vdev_children
; c
++)
1365 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1366 done
, private, flags
));
1373 zio_trim(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1374 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1375 zio_flag_t flags
, enum trim_flag trim_flags
)
1379 ASSERT0(vd
->vdev_children
);
1380 ASSERT0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
1381 ASSERT0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
1382 ASSERT3U(size
, !=, 0);
1384 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, NULL
, size
, size
, done
,
1385 private, ZIO_TYPE_TRIM
, priority
, flags
| ZIO_FLAG_PHYSICAL
,
1386 vd
, offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_TRIM_PIPELINE
);
1387 zio
->io_trim_flags
= trim_flags
;
1393 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1394 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1395 zio_priority_t priority
, zio_flag_t flags
, boolean_t labels
)
1399 ASSERT(vd
->vdev_children
== 0);
1400 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1401 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1402 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1404 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1405 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1406 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1408 zio
->io_prop
.zp_checksum
= checksum
;
1414 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1415 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1416 zio_priority_t priority
, zio_flag_t flags
, boolean_t labels
)
1420 ASSERT(vd
->vdev_children
== 0);
1421 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1422 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1423 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1425 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1426 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1427 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1429 zio
->io_prop
.zp_checksum
= checksum
;
1431 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1433 * zec checksums are necessarily destructive -- they modify
1434 * the end of the write buffer to hold the verifier/checksum.
1435 * Therefore, we must make a local copy in case the data is
1436 * being written to multiple places in parallel.
1438 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1439 abd_copy(wbuf
, data
, size
);
1441 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1448 * Create a child I/O to do some work for us.
1451 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1452 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1453 zio_flag_t flags
, zio_done_func_t
*done
, void *private)
1455 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1459 * vdev child I/Os do not propagate their error to the parent.
1460 * Therefore, for correct operation the caller *must* check for
1461 * and handle the error in the child i/o's done callback.
1462 * The only exceptions are i/os that we don't care about
1463 * (OPTIONAL or REPAIR).
1465 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1468 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1470 * If we have the bp, then the child should perform the
1471 * checksum and the parent need not. This pushes error
1472 * detection as close to the leaves as possible and
1473 * eliminates redundant checksums in the interior nodes.
1475 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1476 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1479 if (vd
->vdev_ops
->vdev_op_leaf
) {
1480 ASSERT0(vd
->vdev_children
);
1481 offset
+= VDEV_LABEL_START_SIZE
;
1484 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1487 * If we've decided to do a repair, the write is not speculative --
1488 * even if the original read was.
1490 if (flags
& ZIO_FLAG_IO_REPAIR
)
1491 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1494 * If we're creating a child I/O that is not associated with a
1495 * top-level vdev, then the child zio is not an allocating I/O.
1496 * If this is a retried I/O then we ignore it since we will
1497 * have already processed the original allocating I/O.
1499 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1500 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1501 ASSERT(pio
->io_metaslab_class
!= NULL
);
1502 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1503 ASSERT(type
== ZIO_TYPE_WRITE
);
1504 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1505 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1506 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1507 pio
->io_child_type
== ZIO_CHILD_GANG
);
1509 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1512 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1513 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1514 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1515 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1521 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1522 zio_type_t type
, zio_priority_t priority
, zio_flag_t flags
,
1523 zio_done_func_t
*done
, void *private)
1527 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1529 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1530 data
, size
, size
, done
, private, type
, priority
,
1531 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1533 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1539 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1541 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1543 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1547 zio_shrink(zio_t
*zio
, uint64_t size
)
1549 ASSERT3P(zio
->io_executor
, ==, NULL
);
1550 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1551 ASSERT3U(size
, <=, zio
->io_size
);
1554 * We don't shrink for raidz because of problems with the
1555 * reconstruction when reading back less than the block size.
1556 * Note, BP_IS_RAIDZ() assumes no compression.
1558 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1559 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1560 /* we are not doing a raw write */
1561 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1562 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1567 * ==========================================================================
1568 * Prepare to read and write logical blocks
1569 * ==========================================================================
1573 zio_read_bp_init(zio_t
*zio
)
1575 blkptr_t
*bp
= zio
->io_bp
;
1577 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1579 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1581 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1582 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1583 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1584 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1585 psize
, psize
, zio_decompress
);
1588 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1589 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1590 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1591 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1592 psize
, psize
, zio_decrypt
);
1595 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1596 int psize
= BPE_GET_PSIZE(bp
);
1597 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1599 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1600 decode_embedded_bp_compressed(bp
, data
);
1601 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1603 ASSERT(!BP_IS_EMBEDDED(bp
));
1604 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1607 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1608 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1614 zio_write_bp_init(zio_t
*zio
)
1616 if (!IO_IS_ALLOCATING(zio
))
1619 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1621 if (zio
->io_bp_override
) {
1622 blkptr_t
*bp
= zio
->io_bp
;
1623 zio_prop_t
*zp
= &zio
->io_prop
;
1625 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1627 *bp
= *zio
->io_bp_override
;
1628 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1630 if (zp
->zp_brtwrite
)
1633 ASSERT(!BP_GET_DEDUP(zio
->io_bp_override
));
1635 if (BP_IS_EMBEDDED(bp
))
1639 * If we've been overridden and nopwrite is set then
1640 * set the flag accordingly to indicate that a nopwrite
1641 * has already occurred.
1643 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1644 ASSERT(!zp
->zp_dedup
);
1645 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1646 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1650 ASSERT(!zp
->zp_nopwrite
);
1652 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1655 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1656 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1658 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1660 BP_SET_DEDUP(bp
, 1);
1661 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1666 * We were unable to handle this as an override bp, treat
1667 * it as a regular write I/O.
1669 zio
->io_bp_override
= NULL
;
1670 *bp
= zio
->io_bp_orig
;
1671 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1678 zio_write_compress(zio_t
*zio
)
1680 spa_t
*spa
= zio
->io_spa
;
1681 zio_prop_t
*zp
= &zio
->io_prop
;
1682 enum zio_compress compress
= zp
->zp_compress
;
1683 blkptr_t
*bp
= zio
->io_bp
;
1684 uint64_t lsize
= zio
->io_lsize
;
1685 uint64_t psize
= zio
->io_size
;
1689 * If our children haven't all reached the ready stage,
1690 * wait for them and then repeat this pipeline stage.
1692 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1693 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1697 if (!IO_IS_ALLOCATING(zio
))
1700 if (zio
->io_children_ready
!= NULL
) {
1702 * Now that all our children are ready, run the callback
1703 * associated with this zio in case it wants to modify the
1704 * data to be written.
1706 ASSERT3U(zp
->zp_level
, >, 0);
1707 zio
->io_children_ready(zio
);
1710 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1711 ASSERT(zio
->io_bp_override
== NULL
);
1713 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1715 * We're rewriting an existing block, which means we're
1716 * working on behalf of spa_sync(). For spa_sync() to
1717 * converge, it must eventually be the case that we don't
1718 * have to allocate new blocks. But compression changes
1719 * the blocksize, which forces a reallocate, and makes
1720 * convergence take longer. Therefore, after the first
1721 * few passes, stop compressing to ensure convergence.
1723 pass
= spa_sync_pass(spa
);
1725 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1726 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1727 ASSERT(!BP_GET_DEDUP(bp
));
1729 if (pass
>= zfs_sync_pass_dont_compress
)
1730 compress
= ZIO_COMPRESS_OFF
;
1732 /* Make sure someone doesn't change their mind on overwrites */
1733 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1734 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1737 /* If it's a compressed write that is not raw, compress the buffer. */
1738 if (compress
!= ZIO_COMPRESS_OFF
&&
1739 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1741 psize
= zio_compress_data(compress
, zio
->io_abd
, &cbuf
, lsize
,
1744 compress
= ZIO_COMPRESS_OFF
;
1745 } else if (psize
>= lsize
) {
1746 compress
= ZIO_COMPRESS_OFF
;
1748 zio_buf_free(cbuf
, lsize
);
1749 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1750 psize
<= BPE_PAYLOAD_SIZE
&&
1751 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1752 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1753 encode_embedded_bp_compressed(bp
,
1754 cbuf
, compress
, lsize
, psize
);
1755 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1756 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1757 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1758 zio_buf_free(cbuf
, lsize
);
1759 bp
->blk_birth
= zio
->io_txg
;
1760 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1761 ASSERT(spa_feature_is_active(spa
,
1762 SPA_FEATURE_EMBEDDED_DATA
));
1766 * Round compressed size up to the minimum allocation
1767 * size of the smallest-ashift device, and zero the
1768 * tail. This ensures that the compressed size of the
1769 * BP (and thus compressratio property) are correct,
1770 * in that we charge for the padding used to fill out
1773 ASSERT3U(spa
->spa_min_alloc
, >=, SPA_MINBLOCKSHIFT
);
1774 size_t rounded
= (size_t)roundup(psize
,
1775 spa
->spa_min_alloc
);
1776 if (rounded
>= lsize
) {
1777 compress
= ZIO_COMPRESS_OFF
;
1778 zio_buf_free(cbuf
, lsize
);
1781 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1782 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1783 abd_zero_off(cdata
, psize
, rounded
- psize
);
1785 zio_push_transform(zio
, cdata
,
1786 psize
, lsize
, NULL
);
1791 * We were unable to handle this as an override bp, treat
1792 * it as a regular write I/O.
1794 zio
->io_bp_override
= NULL
;
1795 *bp
= zio
->io_bp_orig
;
1796 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1798 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1799 zp
->zp_type
== DMU_OT_DNODE
) {
1801 * The DMU actually relies on the zio layer's compression
1802 * to free metadnode blocks that have had all contained
1803 * dnodes freed. As a result, even when doing a raw
1804 * receive, we must check whether the block can be compressed
1807 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1808 zio
->io_abd
, NULL
, lsize
, zp
->zp_complevel
);
1809 if (psize
== 0 || psize
>= lsize
)
1810 compress
= ZIO_COMPRESS_OFF
;
1811 } else if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
&&
1812 !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) {
1814 * If we are raw receiving an encrypted dataset we should not
1815 * take this codepath because it will change the on-disk block
1816 * and decryption will fail.
1818 size_t rounded
= MIN((size_t)roundup(psize
,
1819 spa
->spa_min_alloc
), lsize
);
1821 if (rounded
!= psize
) {
1822 abd_t
*cdata
= abd_alloc_linear(rounded
, B_TRUE
);
1823 abd_zero_off(cdata
, psize
, rounded
- psize
);
1824 abd_copy_off(cdata
, zio
->io_abd
, 0, 0, psize
);
1826 zio_push_transform(zio
, cdata
,
1827 psize
, rounded
, NULL
);
1830 ASSERT3U(psize
, !=, 0);
1834 * The final pass of spa_sync() must be all rewrites, but the first
1835 * few passes offer a trade-off: allocating blocks defers convergence,
1836 * but newly allocated blocks are sequential, so they can be written
1837 * to disk faster. Therefore, we allow the first few passes of
1838 * spa_sync() to allocate new blocks, but force rewrites after that.
1839 * There should only be a handful of blocks after pass 1 in any case.
1841 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1842 BP_GET_PSIZE(bp
) == psize
&&
1843 pass
>= zfs_sync_pass_rewrite
) {
1844 VERIFY3U(psize
, !=, 0);
1845 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1847 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1848 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1851 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1855 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1856 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1857 BP_SET_LSIZE(bp
, lsize
);
1858 BP_SET_TYPE(bp
, zp
->zp_type
);
1859 BP_SET_LEVEL(bp
, zp
->zp_level
);
1860 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1862 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1864 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1865 BP_SET_LSIZE(bp
, lsize
);
1866 BP_SET_TYPE(bp
, zp
->zp_type
);
1867 BP_SET_LEVEL(bp
, zp
->zp_level
);
1868 BP_SET_PSIZE(bp
, psize
);
1869 BP_SET_COMPRESS(bp
, compress
);
1870 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1871 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1872 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1874 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1875 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1876 ASSERT(!zp
->zp_encrypt
||
1877 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1878 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1880 if (zp
->zp_nopwrite
) {
1881 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1882 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1883 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1890 zio_free_bp_init(zio_t
*zio
)
1892 blkptr_t
*bp
= zio
->io_bp
;
1894 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1895 if (BP_GET_DEDUP(bp
))
1896 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1899 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1905 * ==========================================================================
1906 * Execute the I/O pipeline
1907 * ==========================================================================
1911 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1913 spa_t
*spa
= zio
->io_spa
;
1914 zio_type_t t
= zio
->io_type
;
1915 int flags
= (cutinline
? TQ_FRONT
: 0);
1918 * If we're a config writer or a probe, the normal issue and
1919 * interrupt threads may all be blocked waiting for the config lock.
1920 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1922 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1926 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1928 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1932 * If this is a high priority I/O, then use the high priority taskq if
1935 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1936 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1937 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1940 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1943 * NB: We are assuming that the zio can only be dispatched
1944 * to a single taskq at a time. It would be a grievous error
1945 * to dispatch the zio to another taskq at the same time.
1947 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1948 spa_taskq_dispatch_ent(spa
, t
, q
, zio_execute
, zio
, flags
,
1953 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1955 spa_t
*spa
= zio
->io_spa
;
1957 taskq_t
*tq
= taskq_of_curthread();
1959 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1960 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1962 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1963 if (tqs
->stqs_taskq
[i
] == tq
)
1972 zio_issue_async(zio_t
*zio
)
1974 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1980 zio_interrupt(void *zio
)
1982 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1986 zio_delay_interrupt(zio_t
*zio
)
1989 * The timeout_generic() function isn't defined in userspace, so
1990 * rather than trying to implement the function, the zio delay
1991 * functionality has been disabled for userspace builds.
1996 * If io_target_timestamp is zero, then no delay has been registered
1997 * for this IO, thus jump to the end of this function and "skip" the
1998 * delay; issuing it directly to the zio layer.
2000 if (zio
->io_target_timestamp
!= 0) {
2001 hrtime_t now
= gethrtime();
2003 if (now
>= zio
->io_target_timestamp
) {
2005 * This IO has already taken longer than the target
2006 * delay to complete, so we don't want to delay it
2007 * any longer; we "miss" the delay and issue it
2008 * directly to the zio layer. This is likely due to
2009 * the target latency being set to a value less than
2010 * the underlying hardware can satisfy (e.g. delay
2011 * set to 1ms, but the disks take 10ms to complete an
2015 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
2021 hrtime_t diff
= zio
->io_target_timestamp
- now
;
2022 clock_t expire_at_tick
= ddi_get_lbolt() +
2025 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
2026 hrtime_t
, now
, hrtime_t
, diff
);
2028 if (NSEC_TO_TICK(diff
) == 0) {
2029 /* Our delay is less than a jiffy - just spin */
2030 zfs_sleep_until(zio
->io_target_timestamp
);
2034 * Use taskq_dispatch_delay() in the place of
2035 * OpenZFS's timeout_generic().
2037 tid
= taskq_dispatch_delay(system_taskq
,
2038 zio_interrupt
, zio
, TQ_NOSLEEP
,
2040 if (tid
== TASKQID_INVALID
) {
2042 * Couldn't allocate a task. Just
2043 * finish the zio without a delay.
2052 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
2057 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
2059 zio_t
*cio
, *cio_next
;
2060 zio_link_t
*zl
= NULL
;
2061 vdev_t
*vd
= pio
->io_vd
;
2063 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
2064 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
2065 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
2066 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
2067 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
2069 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2070 "delta=%llu queued=%llu io=%llu "
2072 "last=%llu type=%d "
2073 "priority=%d flags=0x%llx stage=0x%x "
2074 "pipeline=0x%x pipeline-trace=0x%x "
2075 "objset=%llu object=%llu "
2076 "level=%llu blkid=%llu "
2077 "offset=%llu size=%llu "
2079 ziodepth
, pio
, pio
->io_timestamp
,
2080 (u_longlong_t
)delta
, pio
->io_delta
, pio
->io_delay
,
2081 vd
? vd
->vdev_path
: "NULL",
2082 vq
? vq
->vq_io_complete_ts
: 0, pio
->io_type
,
2083 pio
->io_priority
, (u_longlong_t
)pio
->io_flags
,
2084 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
2085 (u_longlong_t
)zb
->zb_objset
, (u_longlong_t
)zb
->zb_object
,
2086 (u_longlong_t
)zb
->zb_level
, (u_longlong_t
)zb
->zb_blkid
,
2087 (u_longlong_t
)pio
->io_offset
, (u_longlong_t
)pio
->io_size
,
2089 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
2090 pio
->io_spa
, vd
, zb
, pio
, 0);
2092 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
2093 taskq_empty_ent(&pio
->io_tqent
)) {
2098 mutex_enter(&pio
->io_lock
);
2099 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2100 cio_next
= zio_walk_children(pio
, &zl
);
2101 zio_deadman_impl(cio
, ziodepth
+ 1);
2103 mutex_exit(&pio
->io_lock
);
2107 * Log the critical information describing this zio and all of its children
2108 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2111 zio_deadman(zio_t
*pio
, const char *tag
)
2113 spa_t
*spa
= pio
->io_spa
;
2114 char *name
= spa_name(spa
);
2116 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
2119 zio_deadman_impl(pio
, 0);
2121 switch (spa_get_deadman_failmode(spa
)) {
2122 case ZIO_FAILURE_MODE_WAIT
:
2123 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
2126 case ZIO_FAILURE_MODE_CONTINUE
:
2127 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
2130 case ZIO_FAILURE_MODE_PANIC
:
2131 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
2137 * Execute the I/O pipeline until one of the following occurs:
2138 * (1) the I/O completes; (2) the pipeline stalls waiting for
2139 * dependent child I/Os; (3) the I/O issues, so we're waiting
2140 * for an I/O completion interrupt; (4) the I/O is delegated by
2141 * vdev-level caching or aggregation; (5) the I/O is deferred
2142 * due to vdev-level queueing; (6) the I/O is handed off to
2143 * another thread. In all cases, the pipeline stops whenever
2144 * there's no CPU work; it never burns a thread in cv_wait_io().
2146 * There's no locking on io_stage because there's no legitimate way
2147 * for multiple threads to be attempting to process the same I/O.
2149 static zio_pipe_stage_t
*zio_pipeline
[];
2152 * zio_execute() is a wrapper around the static function
2153 * __zio_execute() so that we can force __zio_execute() to be
2154 * inlined. This reduces stack overhead which is important
2155 * because __zio_execute() is called recursively in several zio
2156 * code paths. zio_execute() itself cannot be inlined because
2157 * it is externally visible.
2160 zio_execute(void *zio
)
2162 fstrans_cookie_t cookie
;
2164 cookie
= spl_fstrans_mark();
2166 spl_fstrans_unmark(cookie
);
2170 * Used to determine if in the current context the stack is sized large
2171 * enough to allow zio_execute() to be called recursively. A minimum
2172 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2175 zio_execute_stack_check(zio_t
*zio
)
2177 #if !defined(HAVE_LARGE_STACKS)
2178 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
2180 /* Executing in txg_sync_thread() context. */
2181 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
2184 /* Pool initialization outside of zio_taskq context. */
2185 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
2186 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
2187 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
2191 #endif /* HAVE_LARGE_STACKS */
2196 __attribute__((always_inline
))
2198 __zio_execute(zio_t
*zio
)
2200 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2202 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
2203 enum zio_stage pipeline
= zio
->io_pipeline
;
2204 enum zio_stage stage
= zio
->io_stage
;
2206 zio
->io_executor
= curthread
;
2208 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
2209 ASSERT(ISP2(stage
));
2210 ASSERT(zio
->io_stall
== NULL
);
2214 } while ((stage
& pipeline
) == 0);
2216 ASSERT(stage
<= ZIO_STAGE_DONE
);
2219 * If we are in interrupt context and this pipeline stage
2220 * will grab a config lock that is held across I/O,
2221 * or may wait for an I/O that needs an interrupt thread
2222 * to complete, issue async to avoid deadlock.
2224 * For VDEV_IO_START, we cut in line so that the io will
2225 * be sent to disk promptly.
2227 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
2228 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
2229 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2230 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2231 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2236 * If the current context doesn't have large enough stacks
2237 * the zio must be issued asynchronously to prevent overflow.
2239 if (zio_execute_stack_check(zio
)) {
2240 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2241 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2242 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2246 zio
->io_stage
= stage
;
2247 zio
->io_pipeline_trace
|= zio
->io_stage
;
2250 * The zio pipeline stage returns the next zio to execute
2251 * (typically the same as this one), or NULL if we should
2254 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2263 * ==========================================================================
2264 * Initiate I/O, either sync or async
2265 * ==========================================================================
2268 zio_wait(zio_t
*zio
)
2271 * Some routines, like zio_free_sync(), may return a NULL zio
2272 * to avoid the performance overhead of creating and then destroying
2273 * an unneeded zio. For the callers' simplicity, we accept a NULL
2274 * zio and ignore it.
2279 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2282 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2283 ASSERT3P(zio
->io_executor
, ==, NULL
);
2285 zio
->io_waiter
= curthread
;
2286 ASSERT0(zio
->io_queued_timestamp
);
2287 zio
->io_queued_timestamp
= gethrtime();
2291 mutex_enter(&zio
->io_lock
);
2292 while (zio
->io_executor
!= NULL
) {
2293 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2294 ddi_get_lbolt() + timeout
);
2296 if (zfs_deadman_enabled
&& error
== -1 &&
2297 gethrtime() - zio
->io_queued_timestamp
>
2298 spa_deadman_ziotime(zio
->io_spa
)) {
2299 mutex_exit(&zio
->io_lock
);
2300 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2301 zio_deadman(zio
, FTAG
);
2302 mutex_enter(&zio
->io_lock
);
2305 mutex_exit(&zio
->io_lock
);
2307 error
= zio
->io_error
;
2314 zio_nowait(zio_t
*zio
)
2317 * See comment in zio_wait().
2322 ASSERT3P(zio
->io_executor
, ==, NULL
);
2324 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2325 list_is_empty(&zio
->io_parent_list
)) {
2329 * This is a logical async I/O with no parent to wait for it.
2330 * We add it to the spa_async_root_zio "Godfather" I/O which
2331 * will ensure they complete prior to unloading the pool.
2333 spa_t
*spa
= zio
->io_spa
;
2334 pio
= spa
->spa_async_zio_root
[CPU_SEQID_UNSTABLE
];
2336 zio_add_child(pio
, zio
);
2339 ASSERT0(zio
->io_queued_timestamp
);
2340 zio
->io_queued_timestamp
= gethrtime();
2345 * ==========================================================================
2346 * Reexecute, cancel, or suspend/resume failed I/O
2347 * ==========================================================================
2351 zio_reexecute(void *arg
)
2354 zio_t
*cio
, *cio_next
;
2356 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2357 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2358 ASSERT(pio
->io_gang_leader
== NULL
);
2359 ASSERT(pio
->io_gang_tree
== NULL
);
2361 pio
->io_flags
= pio
->io_orig_flags
;
2362 pio
->io_stage
= pio
->io_orig_stage
;
2363 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2364 pio
->io_reexecute
= 0;
2365 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2366 pio
->io_pipeline_trace
= 0;
2368 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2369 pio
->io_state
[w
] = 0;
2370 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2371 pio
->io_child_error
[c
] = 0;
2373 if (IO_IS_ALLOCATING(pio
))
2374 BP_ZERO(pio
->io_bp
);
2377 * As we reexecute pio's children, new children could be created.
2378 * New children go to the head of pio's io_child_list, however,
2379 * so we will (correctly) not reexecute them. The key is that
2380 * the remainder of pio's io_child_list, from 'cio_next' onward,
2381 * cannot be affected by any side effects of reexecuting 'cio'.
2383 zio_link_t
*zl
= NULL
;
2384 mutex_enter(&pio
->io_lock
);
2385 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2386 cio_next
= zio_walk_children(pio
, &zl
);
2387 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2388 pio
->io_children
[cio
->io_child_type
][w
]++;
2389 mutex_exit(&pio
->io_lock
);
2391 mutex_enter(&pio
->io_lock
);
2393 mutex_exit(&pio
->io_lock
);
2396 * Now that all children have been reexecuted, execute the parent.
2397 * We don't reexecute "The Godfather" I/O here as it's the
2398 * responsibility of the caller to wait on it.
2400 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2401 pio
->io_queued_timestamp
= gethrtime();
2407 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2409 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2410 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2411 "failure and the failure mode property for this pool "
2412 "is set to panic.", spa_name(spa
));
2414 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2415 "failure and has been suspended.\n", spa_name(spa
));
2417 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2420 mutex_enter(&spa
->spa_suspend_lock
);
2422 if (spa
->spa_suspend_zio_root
== NULL
)
2423 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2424 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2425 ZIO_FLAG_GODFATHER
);
2427 spa
->spa_suspended
= reason
;
2430 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2431 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2432 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2433 ASSERT(zio_unique_parent(zio
) == NULL
);
2434 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2435 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2438 mutex_exit(&spa
->spa_suspend_lock
);
2442 zio_resume(spa_t
*spa
)
2447 * Reexecute all previously suspended i/o.
2449 mutex_enter(&spa
->spa_suspend_lock
);
2450 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2451 cv_broadcast(&spa
->spa_suspend_cv
);
2452 pio
= spa
->spa_suspend_zio_root
;
2453 spa
->spa_suspend_zio_root
= NULL
;
2454 mutex_exit(&spa
->spa_suspend_lock
);
2460 return (zio_wait(pio
));
2464 zio_resume_wait(spa_t
*spa
)
2466 mutex_enter(&spa
->spa_suspend_lock
);
2467 while (spa_suspended(spa
))
2468 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2469 mutex_exit(&spa
->spa_suspend_lock
);
2473 * ==========================================================================
2476 * A gang block is a collection of small blocks that looks to the DMU
2477 * like one large block. When zio_dva_allocate() cannot find a block
2478 * of the requested size, due to either severe fragmentation or the pool
2479 * being nearly full, it calls zio_write_gang_block() to construct the
2480 * block from smaller fragments.
2482 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2483 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2484 * an indirect block: it's an array of block pointers. It consumes
2485 * only one sector and hence is allocatable regardless of fragmentation.
2486 * The gang header's bps point to its gang members, which hold the data.
2488 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2489 * as the verifier to ensure uniqueness of the SHA256 checksum.
2490 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2491 * not the gang header. This ensures that data block signatures (needed for
2492 * deduplication) are independent of how the block is physically stored.
2494 * Gang blocks can be nested: a gang member may itself be a gang block.
2495 * Thus every gang block is a tree in which root and all interior nodes are
2496 * gang headers, and the leaves are normal blocks that contain user data.
2497 * The root of the gang tree is called the gang leader.
2499 * To perform any operation (read, rewrite, free, claim) on a gang block,
2500 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2501 * in the io_gang_tree field of the original logical i/o by recursively
2502 * reading the gang leader and all gang headers below it. This yields
2503 * an in-core tree containing the contents of every gang header and the
2504 * bps for every constituent of the gang block.
2506 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2507 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2508 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2509 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2510 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2511 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2512 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2513 * of the gang header plus zio_checksum_compute() of the data to update the
2514 * gang header's blk_cksum as described above.
2516 * The two-phase assemble/issue model solves the problem of partial failure --
2517 * what if you'd freed part of a gang block but then couldn't read the
2518 * gang header for another part? Assembling the entire gang tree first
2519 * ensures that all the necessary gang header I/O has succeeded before
2520 * starting the actual work of free, claim, or write. Once the gang tree
2521 * is assembled, free and claim are in-memory operations that cannot fail.
2523 * In the event that a gang write fails, zio_dva_unallocate() walks the
2524 * gang tree to immediately free (i.e. insert back into the space map)
2525 * everything we've allocated. This ensures that we don't get ENOSPC
2526 * errors during repeated suspend/resume cycles due to a flaky device.
2528 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2529 * the gang tree, we won't modify the block, so we can safely defer the free
2530 * (knowing that the block is still intact). If we *can* assemble the gang
2531 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2532 * each constituent bp and we can allocate a new block on the next sync pass.
2534 * In all cases, the gang tree allows complete recovery from partial failure.
2535 * ==========================================================================
2539 zio_gang_issue_func_done(zio_t
*zio
)
2541 abd_free(zio
->io_abd
);
2545 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2551 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2552 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2553 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2554 &pio
->io_bookmark
));
2558 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2565 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2566 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2567 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2568 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2571 * As we rewrite each gang header, the pipeline will compute
2572 * a new gang block header checksum for it; but no one will
2573 * compute a new data checksum, so we do that here. The one
2574 * exception is the gang leader: the pipeline already computed
2575 * its data checksum because that stage precedes gang assembly.
2576 * (Presently, nothing actually uses interior data checksums;
2577 * this is just good hygiene.)
2579 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2580 abd_t
*buf
= abd_get_offset(data
, offset
);
2582 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2583 buf
, BP_GET_PSIZE(bp
));
2588 * If we are here to damage data for testing purposes,
2589 * leave the GBH alone so that we can detect the damage.
2591 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2592 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2594 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2595 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2596 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2597 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2604 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2607 (void) gn
, (void) data
, (void) offset
;
2609 zio_t
*zio
= zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2610 ZIO_GANG_CHILD_FLAGS(pio
));
2612 zio
= zio_null(pio
, pio
->io_spa
,
2613 NULL
, NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
));
2619 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2622 (void) gn
, (void) data
, (void) offset
;
2623 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2624 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2627 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2636 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2638 static zio_gang_node_t
*
2639 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2641 zio_gang_node_t
*gn
;
2643 ASSERT(*gnpp
== NULL
);
2645 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2646 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2653 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2655 zio_gang_node_t
*gn
= *gnpp
;
2657 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2658 ASSERT(gn
->gn_child
[g
] == NULL
);
2660 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2661 kmem_free(gn
, sizeof (*gn
));
2666 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2668 zio_gang_node_t
*gn
= *gnpp
;
2673 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2674 zio_gang_tree_free(&gn
->gn_child
[g
]);
2676 zio_gang_node_free(gnpp
);
2680 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2682 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2683 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2685 ASSERT(gio
->io_gang_leader
== gio
);
2686 ASSERT(BP_IS_GANG(bp
));
2688 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2689 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2690 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2694 zio_gang_tree_assemble_done(zio_t
*zio
)
2696 zio_t
*gio
= zio
->io_gang_leader
;
2697 zio_gang_node_t
*gn
= zio
->io_private
;
2698 blkptr_t
*bp
= zio
->io_bp
;
2700 ASSERT(gio
== zio_unique_parent(zio
));
2701 ASSERT(list_is_empty(&zio
->io_child_list
));
2706 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2707 if (BP_SHOULD_BYTESWAP(bp
))
2708 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2710 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2711 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2712 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2714 abd_free(zio
->io_abd
);
2716 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2717 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2718 if (!BP_IS_GANG(gbp
))
2720 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2725 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2728 zio_t
*gio
= pio
->io_gang_leader
;
2731 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2732 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2733 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2736 * If you're a gang header, your data is in gn->gn_gbh.
2737 * If you're a gang member, your data is in 'data' and gn == NULL.
2739 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2742 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2744 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2745 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2746 if (BP_IS_HOLE(gbp
))
2748 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2750 offset
+= BP_GET_PSIZE(gbp
);
2754 if (gn
== gio
->io_gang_tree
)
2755 ASSERT3U(gio
->io_size
, ==, offset
);
2762 zio_gang_assemble(zio_t
*zio
)
2764 blkptr_t
*bp
= zio
->io_bp
;
2766 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2767 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2769 zio
->io_gang_leader
= zio
;
2771 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2777 zio_gang_issue(zio_t
*zio
)
2779 blkptr_t
*bp
= zio
->io_bp
;
2781 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2785 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2786 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2788 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2789 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2792 zio_gang_tree_free(&zio
->io_gang_tree
);
2794 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2800 zio_write_gang_member_ready(zio_t
*zio
)
2802 zio_t
*pio
= zio_unique_parent(zio
);
2803 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2804 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2806 zio_t
*gio __maybe_unused
= zio
->io_gang_leader
;
2808 if (BP_IS_HOLE(zio
->io_bp
))
2811 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2813 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2814 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2815 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2816 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2817 VERIFY3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2819 mutex_enter(&pio
->io_lock
);
2820 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2821 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2822 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2823 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2824 DVA_SET_ASIZE(&pdva
[d
], asize
);
2826 mutex_exit(&pio
->io_lock
);
2830 zio_write_gang_done(zio_t
*zio
)
2833 * The io_abd field will be NULL for a zio with no data. The io_flags
2834 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2835 * check for it here as it is cleared in zio_ready.
2837 if (zio
->io_abd
!= NULL
)
2838 abd_free(zio
->io_abd
);
2842 zio_write_gang_block(zio_t
*pio
, metaslab_class_t
*mc
)
2844 spa_t
*spa
= pio
->io_spa
;
2845 blkptr_t
*bp
= pio
->io_bp
;
2846 zio_t
*gio
= pio
->io_gang_leader
;
2848 zio_gang_node_t
*gn
, **gnpp
;
2849 zio_gbh_phys_t
*gbh
;
2851 uint64_t txg
= pio
->io_txg
;
2852 uint64_t resid
= pio
->io_size
;
2854 int copies
= gio
->io_prop
.zp_copies
;
2857 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2860 * If one copy was requested, store 2 copies of the GBH, so that we
2861 * can still traverse all the data (e.g. to free or scrub) even if a
2862 * block is damaged. Note that we can't store 3 copies of the GBH in
2863 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
2865 int gbh_copies
= copies
;
2866 if (gbh_copies
== 1) {
2867 gbh_copies
= MIN(2, spa_max_replication(spa
));
2870 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2871 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2872 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2875 flags
|= METASLAB_ASYNC_ALLOC
;
2876 VERIFY(zfs_refcount_held(&mc
->mc_allocator
[pio
->io_allocator
].
2877 mca_alloc_slots
, pio
));
2880 * The logical zio has already placed a reservation for
2881 * 'copies' allocation slots but gang blocks may require
2882 * additional copies. These additional copies
2883 * (i.e. gbh_copies - copies) are guaranteed to succeed
2884 * since metaslab_class_throttle_reserve() always allows
2885 * additional reservations for gang blocks.
2887 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2888 pio
->io_allocator
, pio
, flags
));
2891 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2892 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2893 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2895 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2896 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2900 * If we failed to allocate the gang block header then
2901 * we remove any additional allocation reservations that
2902 * we placed here. The original reservation will
2903 * be removed when the logical I/O goes to the ready
2906 metaslab_class_throttle_unreserve(mc
,
2907 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2910 pio
->io_error
= error
;
2915 gnpp
= &gio
->io_gang_tree
;
2917 gnpp
= pio
->io_private
;
2918 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2921 gn
= zio_gang_node_alloc(gnpp
);
2923 memset(gbh
, 0, SPA_GANGBLOCKSIZE
);
2924 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2927 * Create the gang header.
2929 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2930 zio_write_gang_done
, NULL
, pio
->io_priority
,
2931 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2934 * Create and nowait the gang children.
2936 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2937 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2939 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2941 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2942 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2943 zp
.zp_complevel
= gio
->io_prop
.zp_complevel
;
2944 zp
.zp_type
= DMU_OT_NONE
;
2946 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2947 zp
.zp_dedup
= B_FALSE
;
2948 zp
.zp_dedup_verify
= B_FALSE
;
2949 zp
.zp_nopwrite
= B_FALSE
;
2950 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2951 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2952 memset(zp
.zp_salt
, 0, ZIO_DATA_SALT_LEN
);
2953 memset(zp
.zp_iv
, 0, ZIO_DATA_IV_LEN
);
2954 memset(zp
.zp_mac
, 0, ZIO_DATA_MAC_LEN
);
2956 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2957 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2958 resid
) : NULL
, lsize
, lsize
, &zp
,
2959 zio_write_gang_member_ready
, NULL
,
2960 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2961 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2963 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2964 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2968 * Gang children won't throttle but we should
2969 * account for their work, so reserve an allocation
2970 * slot for them here.
2972 VERIFY(metaslab_class_throttle_reserve(mc
,
2973 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2979 * Set pio's pipeline to just wait for zio to finish.
2981 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2984 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2986 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2994 * The zio_nop_write stage in the pipeline determines if allocating a
2995 * new bp is necessary. The nopwrite feature can handle writes in
2996 * either syncing or open context (i.e. zil writes) and as a result is
2997 * mutually exclusive with dedup.
2999 * By leveraging a cryptographically secure checksum, such as SHA256, we
3000 * can compare the checksums of the new data and the old to determine if
3001 * allocating a new block is required. Note that our requirements for
3002 * cryptographic strength are fairly weak: there can't be any accidental
3003 * hash collisions, but we don't need to be secure against intentional
3004 * (malicious) collisions. To trigger a nopwrite, you have to be able
3005 * to write the file to begin with, and triggering an incorrect (hash
3006 * collision) nopwrite is no worse than simply writing to the file.
3007 * That said, there are no known attacks against the checksum algorithms
3008 * used for nopwrite, assuming that the salt and the checksums
3009 * themselves remain secret.
3012 zio_nop_write(zio_t
*zio
)
3014 blkptr_t
*bp
= zio
->io_bp
;
3015 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3016 zio_prop_t
*zp
= &zio
->io_prop
;
3018 ASSERT(BP_IS_HOLE(bp
));
3019 ASSERT(BP_GET_LEVEL(bp
) == 0);
3020 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3021 ASSERT(zp
->zp_nopwrite
);
3022 ASSERT(!zp
->zp_dedup
);
3023 ASSERT(zio
->io_bp_override
== NULL
);
3024 ASSERT(IO_IS_ALLOCATING(zio
));
3027 * Check to see if the original bp and the new bp have matching
3028 * characteristics (i.e. same checksum, compression algorithms, etc).
3029 * If they don't then just continue with the pipeline which will
3030 * allocate a new bp.
3032 if (BP_IS_HOLE(bp_orig
) ||
3033 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
3034 ZCHECKSUM_FLAG_NOPWRITE
) ||
3035 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
3036 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
3037 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
3038 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
3039 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
3043 * If the checksums match then reset the pipeline so that we
3044 * avoid allocating a new bp and issuing any I/O.
3046 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
3047 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3048 ZCHECKSUM_FLAG_NOPWRITE
);
3049 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
3050 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
3051 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
3052 ASSERT3U(bp
->blk_prop
, ==, bp_orig
->blk_prop
);
3055 * If we're overwriting a block that is currently on an
3056 * indirect vdev, then ignore the nopwrite request and
3057 * allow a new block to be allocated on a concrete vdev.
3059 spa_config_enter(zio
->io_spa
, SCL_VDEV
, FTAG
, RW_READER
);
3060 for (int d
= 0; d
< BP_GET_NDVAS(bp_orig
); d
++) {
3061 vdev_t
*tvd
= vdev_lookup_top(zio
->io_spa
,
3062 DVA_GET_VDEV(&bp_orig
->blk_dva
[d
]));
3063 if (tvd
->vdev_ops
== &vdev_indirect_ops
) {
3064 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
3068 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
3071 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3072 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
3079 * ==========================================================================
3080 * Block Reference Table
3081 * ==========================================================================
3084 zio_brt_free(zio_t
*zio
)
3090 if (BP_GET_LEVEL(bp
) > 0 ||
3091 BP_IS_METADATA(bp
) ||
3092 !brt_maybe_exists(zio
->io_spa
, bp
)) {
3096 if (!brt_entry_decref(zio
->io_spa
, bp
)) {
3098 * This isn't the last reference, so we cannot free
3101 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3108 * ==========================================================================
3110 * ==========================================================================
3113 zio_ddt_child_read_done(zio_t
*zio
)
3115 blkptr_t
*bp
= zio
->io_bp
;
3116 ddt_entry_t
*dde
= zio
->io_private
;
3118 zio_t
*pio
= zio_unique_parent(zio
);
3120 mutex_enter(&pio
->io_lock
);
3121 ddp
= ddt_phys_select(dde
, bp
);
3122 if (zio
->io_error
== 0)
3123 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
3125 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
3126 dde
->dde_repair_abd
= zio
->io_abd
;
3128 abd_free(zio
->io_abd
);
3129 mutex_exit(&pio
->io_lock
);
3133 zio_ddt_read_start(zio_t
*zio
)
3135 blkptr_t
*bp
= zio
->io_bp
;
3137 ASSERT(BP_GET_DEDUP(bp
));
3138 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3139 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3141 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3142 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3143 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
3144 ddt_phys_t
*ddp
= dde
->dde_phys
;
3145 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
3148 ASSERT(zio
->io_vsd
== NULL
);
3151 if (ddp_self
== NULL
)
3154 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
3155 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
3157 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
3159 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
3160 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
3161 zio
->io_size
, zio_ddt_child_read_done
, dde
,
3162 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
3163 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
3168 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
3169 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
3170 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
3176 zio_ddt_read_done(zio_t
*zio
)
3178 blkptr_t
*bp
= zio
->io_bp
;
3180 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
3184 ASSERT(BP_GET_DEDUP(bp
));
3185 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3186 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3188 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3189 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3190 ddt_entry_t
*dde
= zio
->io_vsd
;
3192 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
3196 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
3197 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
3200 if (dde
->dde_repair_abd
!= NULL
) {
3201 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
3203 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
3205 ddt_repair_done(ddt
, dde
);
3209 ASSERT(zio
->io_vsd
== NULL
);
3215 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
3217 spa_t
*spa
= zio
->io_spa
;
3218 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
3220 ASSERT(!(zio
->io_bp_override
&& do_raw
));
3223 * Note: we compare the original data, not the transformed data,
3224 * because when zio->io_bp is an override bp, we will not have
3225 * pushed the I/O transforms. That's an important optimization
3226 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3227 * However, we should never get a raw, override zio so in these
3228 * cases we can compare the io_abd directly. This is useful because
3229 * it allows us to do dedup verification even if we don't have access
3230 * to the original data (for instance, if the encryption keys aren't
3234 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3235 zio_t
*lio
= dde
->dde_lead_zio
[p
];
3237 if (lio
!= NULL
&& do_raw
) {
3238 return (lio
->io_size
!= zio
->io_size
||
3239 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
3240 } else if (lio
!= NULL
) {
3241 return (lio
->io_orig_size
!= zio
->io_orig_size
||
3242 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
3246 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3247 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3249 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
3250 blkptr_t blk
= *zio
->io_bp
;
3255 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3256 psize
= BP_GET_PSIZE(&blk
);
3258 if (psize
!= zio
->io_size
)
3263 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
3265 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
3266 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
3267 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
3268 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
3271 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
3272 error
= SET_ERROR(ENOENT
);
3277 return (error
!= 0);
3278 } else if (ddp
->ddp_phys_birth
!= 0) {
3279 arc_buf_t
*abuf
= NULL
;
3280 arc_flags_t aflags
= ARC_FLAG_WAIT
;
3281 blkptr_t blk
= *zio
->io_bp
;
3284 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3286 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3291 error
= arc_read(NULL
, spa
, &blk
,
3292 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3293 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3294 &aflags
, &zio
->io_bookmark
);
3297 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3298 zio
->io_orig_size
) != 0)
3299 error
= SET_ERROR(ENOENT
);
3300 arc_buf_destroy(abuf
, &abuf
);
3304 return (error
!= 0);
3312 zio_ddt_child_write_ready(zio_t
*zio
)
3314 int p
= zio
->io_prop
.zp_copies
;
3315 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3316 ddt_entry_t
*dde
= zio
->io_private
;
3317 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3325 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3327 ddt_phys_fill(ddp
, zio
->io_bp
);
3329 zio_link_t
*zl
= NULL
;
3330 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3331 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3337 zio_ddt_child_write_done(zio_t
*zio
)
3339 int p
= zio
->io_prop
.zp_copies
;
3340 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3341 ddt_entry_t
*dde
= zio
->io_private
;
3342 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3346 ASSERT(ddp
->ddp_refcnt
== 0);
3347 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3348 dde
->dde_lead_zio
[p
] = NULL
;
3350 if (zio
->io_error
== 0) {
3351 zio_link_t
*zl
= NULL
;
3352 while (zio_walk_parents(zio
, &zl
) != NULL
)
3353 ddt_phys_addref(ddp
);
3355 ddt_phys_clear(ddp
);
3362 zio_ddt_write(zio_t
*zio
)
3364 spa_t
*spa
= zio
->io_spa
;
3365 blkptr_t
*bp
= zio
->io_bp
;
3366 uint64_t txg
= zio
->io_txg
;
3367 zio_prop_t
*zp
= &zio
->io_prop
;
3368 int p
= zp
->zp_copies
;
3370 ddt_t
*ddt
= ddt_select(spa
, bp
);
3374 ASSERT(BP_GET_DEDUP(bp
));
3375 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3376 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3377 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3380 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3381 ddp
= &dde
->dde_phys
[p
];
3383 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3385 * If we're using a weak checksum, upgrade to a strong checksum
3386 * and try again. If we're already using a strong checksum,
3387 * we can't resolve it, so just convert to an ordinary write.
3388 * (And automatically e-mail a paper to Nature?)
3390 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3391 ZCHECKSUM_FLAG_DEDUP
)) {
3392 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3393 zio_pop_transforms(zio
);
3394 zio
->io_stage
= ZIO_STAGE_OPEN
;
3397 zp
->zp_dedup
= B_FALSE
;
3398 BP_SET_DEDUP(bp
, B_FALSE
);
3400 ASSERT(!BP_GET_DEDUP(bp
));
3401 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3406 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3407 if (ddp
->ddp_phys_birth
!= 0)
3408 ddt_bp_fill(ddp
, bp
, txg
);
3409 if (dde
->dde_lead_zio
[p
] != NULL
)
3410 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3412 ddt_phys_addref(ddp
);
3413 } else if (zio
->io_bp_override
) {
3414 ASSERT(bp
->blk_birth
== txg
);
3415 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3416 ddt_phys_fill(ddp
, bp
);
3417 ddt_phys_addref(ddp
);
3419 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3420 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3421 zio_ddt_child_write_ready
, NULL
,
3422 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3423 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3425 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3426 dde
->dde_lead_zio
[p
] = cio
;
3436 static ddt_entry_t
*freedde
; /* for debugging */
3439 zio_ddt_free(zio_t
*zio
)
3441 spa_t
*spa
= zio
->io_spa
;
3442 blkptr_t
*bp
= zio
->io_bp
;
3443 ddt_t
*ddt
= ddt_select(spa
, bp
);
3447 ASSERT(BP_GET_DEDUP(bp
));
3448 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3451 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3453 ddp
= ddt_phys_select(dde
, bp
);
3455 ddt_phys_decref(ddp
);
3463 * ==========================================================================
3464 * Allocate and free blocks
3465 * ==========================================================================
3469 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3473 ASSERT(MUTEX_HELD(&spa
->spa_allocs
[allocator
].spaa_lock
));
3475 zio
= avl_first(&spa
->spa_allocs
[allocator
].spaa_tree
);
3479 ASSERT(IO_IS_ALLOCATING(zio
));
3482 * Try to place a reservation for this zio. If we're unable to
3483 * reserve then we throttle.
3485 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3486 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3487 zio
->io_prop
.zp_copies
, allocator
, zio
, 0)) {
3491 avl_remove(&spa
->spa_allocs
[allocator
].spaa_tree
, zio
);
3492 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3498 zio_dva_throttle(zio_t
*zio
)
3500 spa_t
*spa
= zio
->io_spa
;
3502 metaslab_class_t
*mc
;
3504 /* locate an appropriate allocation class */
3505 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3506 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3508 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3509 !mc
->mc_alloc_throttle_enabled
||
3510 zio
->io_child_type
== ZIO_CHILD_GANG
||
3511 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3515 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3516 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3517 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3518 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3520 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
3522 * We want to try to use as many allocators as possible to help improve
3523 * performance, but we also want logically adjacent IOs to be physically
3524 * adjacent to improve sequential read performance. We chunk each object
3525 * into 2^20 block regions, and then hash based on the objset, object,
3526 * level, and region to accomplish both of these goals.
3528 int allocator
= (uint_t
)cityhash4(bm
->zb_objset
, bm
->zb_object
,
3529 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
3530 zio
->io_allocator
= allocator
;
3531 zio
->io_metaslab_class
= mc
;
3532 mutex_enter(&spa
->spa_allocs
[allocator
].spaa_lock
);
3533 avl_add(&spa
->spa_allocs
[allocator
].spaa_tree
, zio
);
3534 nio
= zio_io_to_allocate(spa
, allocator
);
3535 mutex_exit(&spa
->spa_allocs
[allocator
].spaa_lock
);
3540 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3544 mutex_enter(&spa
->spa_allocs
[allocator
].spaa_lock
);
3545 zio
= zio_io_to_allocate(spa
, allocator
);
3546 mutex_exit(&spa
->spa_allocs
[allocator
].spaa_lock
);
3550 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3551 ASSERT0(zio
->io_error
);
3552 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3556 zio_dva_allocate(zio_t
*zio
)
3558 spa_t
*spa
= zio
->io_spa
;
3559 metaslab_class_t
*mc
;
3560 blkptr_t
*bp
= zio
->io_bp
;
3564 if (zio
->io_gang_leader
== NULL
) {
3565 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3566 zio
->io_gang_leader
= zio
;
3569 ASSERT(BP_IS_HOLE(bp
));
3570 ASSERT0(BP_GET_NDVAS(bp
));
3571 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3572 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3573 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3575 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3576 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3577 flags
|= METASLAB_DONT_THROTTLE
;
3578 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3579 flags
|= METASLAB_GANG_CHILD
;
3580 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3581 flags
|= METASLAB_ASYNC_ALLOC
;
3584 * if not already chosen, locate an appropriate allocation class
3586 mc
= zio
->io_metaslab_class
;
3588 mc
= spa_preferred_class(spa
, zio
->io_size
,
3589 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3590 zio
->io_prop
.zp_zpl_smallblk
);
3591 zio
->io_metaslab_class
= mc
;
3595 * Try allocating the block in the usual metaslab class.
3596 * If that's full, allocate it in the normal class.
3597 * If that's full, allocate as a gang block,
3598 * and if all are full, the allocation fails (which shouldn't happen).
3600 * Note that we do not fall back on embedded slog (ZIL) space, to
3601 * preserve unfragmented slog space, which is critical for decent
3602 * sync write performance. If a log allocation fails, we will fall
3603 * back to spa_sync() which is abysmal for performance.
3605 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3606 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3607 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3610 * Fallback to normal class when an alloc class is full
3612 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3614 * If throttling, transfer reservation over to normal class.
3615 * The io_allocator slot can remain the same even though we
3616 * are switching classes.
3618 if (mc
->mc_alloc_throttle_enabled
&&
3619 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3620 metaslab_class_throttle_unreserve(mc
,
3621 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3622 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3624 VERIFY(metaslab_class_throttle_reserve(
3625 spa_normal_class(spa
),
3626 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3627 flags
| METASLAB_MUST_RESERVE
));
3629 zio
->io_metaslab_class
= mc
= spa_normal_class(spa
);
3630 if (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
) {
3631 zfs_dbgmsg("%s: metaslab allocation failure, "
3632 "trying normal class: zio %px, size %llu, error %d",
3633 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3637 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3638 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3639 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3642 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
) {
3643 if (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
) {
3644 zfs_dbgmsg("%s: metaslab allocation failure, "
3645 "trying ganging: zio %px, size %llu, error %d",
3646 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3649 return (zio_write_gang_block(zio
, mc
));
3652 if (error
!= ENOSPC
||
3653 (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
)) {
3654 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3655 "size %llu, error %d",
3656 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3659 zio
->io_error
= error
;
3666 zio_dva_free(zio_t
*zio
)
3668 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3674 zio_dva_claim(zio_t
*zio
)
3678 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3680 zio
->io_error
= error
;
3686 * Undo an allocation. This is used by zio_done() when an I/O fails
3687 * and we want to give back the block we just allocated.
3688 * This handles both normal blocks and gang blocks.
3691 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3693 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3694 ASSERT(zio
->io_bp_override
== NULL
);
3696 if (!BP_IS_HOLE(bp
))
3697 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3700 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3701 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3702 &gn
->gn_gbh
->zg_blkptr
[g
]);
3708 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3711 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3712 uint64_t size
, boolean_t
*slog
)
3715 zio_alloc_list_t io_alloc_list
;
3717 ASSERT(txg
> spa_syncing_txg(spa
));
3719 metaslab_trace_init(&io_alloc_list
);
3722 * Block pointer fields are useful to metaslabs for stats and debugging.
3723 * Fill in the obvious ones before calling into metaslab_alloc().
3725 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3726 BP_SET_PSIZE(new_bp
, size
);
3727 BP_SET_LEVEL(new_bp
, 0);
3730 * When allocating a zil block, we don't have information about
3731 * the final destination of the block except the objset it's part
3732 * of, so we just hash the objset ID to pick the allocator to get
3735 int flags
= METASLAB_FASTWRITE
| METASLAB_ZIL
;
3736 int allocator
= (uint_t
)cityhash4(0, 0, 0,
3737 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
;
3738 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3739 txg
, NULL
, flags
, &io_alloc_list
, NULL
, allocator
);
3740 *slog
= (error
== 0);
3742 error
= metaslab_alloc(spa
, spa_embedded_log_class(spa
), size
,
3743 new_bp
, 1, txg
, NULL
, flags
,
3744 &io_alloc_list
, NULL
, allocator
);
3747 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3748 new_bp
, 1, txg
, NULL
, flags
,
3749 &io_alloc_list
, NULL
, allocator
);
3751 metaslab_trace_fini(&io_alloc_list
);
3754 BP_SET_LSIZE(new_bp
, size
);
3755 BP_SET_PSIZE(new_bp
, size
);
3756 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3757 BP_SET_CHECKSUM(new_bp
,
3758 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3759 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3760 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3761 BP_SET_LEVEL(new_bp
, 0);
3762 BP_SET_DEDUP(new_bp
, 0);
3763 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3766 * encrypted blocks will require an IV and salt. We generate
3767 * these now since we will not be rewriting the bp at
3770 if (os
->os_encrypted
) {
3771 uint8_t iv
[ZIO_DATA_IV_LEN
];
3772 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3774 BP_SET_CRYPT(new_bp
, B_TRUE
);
3775 VERIFY0(spa_crypt_get_salt(spa
,
3776 dmu_objset_id(os
), salt
));
3777 VERIFY0(zio_crypt_generate_iv(iv
));
3779 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3782 zfs_dbgmsg("%s: zil block allocation failure: "
3783 "size %llu, error %d", spa_name(spa
), (u_longlong_t
)size
,
3791 * ==========================================================================
3792 * Read and write to physical devices
3793 * ==========================================================================
3797 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3798 * stops after this stage and will resume upon I/O completion.
3799 * However, there are instances where the vdev layer may need to
3800 * continue the pipeline when an I/O was not issued. Since the I/O
3801 * that was sent to the vdev layer might be different than the one
3802 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3803 * force the underlying vdev layers to call either zio_execute() or
3804 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3807 zio_vdev_io_start(zio_t
*zio
)
3809 vdev_t
*vd
= zio
->io_vd
;
3811 spa_t
*spa
= zio
->io_spa
;
3815 ASSERT(zio
->io_error
== 0);
3816 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3819 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3820 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3823 * The mirror_ops handle multiple DVAs in a single BP.
3825 vdev_mirror_ops
.vdev_op_io_start(zio
);
3829 ASSERT3P(zio
->io_logical
, !=, zio
);
3830 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3831 ASSERT(spa
->spa_trust_config
);
3834 * Note: the code can handle other kinds of writes,
3835 * but we don't expect them.
3837 if (zio
->io_vd
->vdev_noalloc
) {
3838 ASSERT(zio
->io_flags
&
3839 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3840 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3844 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3846 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3847 P2PHASE(zio
->io_size
, align
) != 0) {
3848 /* Transform logical writes to be a full physical block size. */
3849 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3850 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3851 ASSERT(vd
== vd
->vdev_top
);
3852 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3853 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3854 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3856 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3860 * If this is not a physical io, make sure that it is properly aligned
3861 * before proceeding.
3863 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3864 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3865 ASSERT0(P2PHASE(zio
->io_size
, align
));
3868 * For physical writes, we allow 512b aligned writes and assume
3869 * the device will perform a read-modify-write as necessary.
3871 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3872 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3875 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3878 * If this is a repair I/O, and there's no self-healing involved --
3879 * that is, we're just resilvering what we expect to resilver --
3880 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3881 * This prevents spurious resilvering.
3883 * There are a few ways that we can end up creating these spurious
3886 * 1. A resilver i/o will be issued if any DVA in the BP has a
3887 * dirty DTL. The mirror code will issue resilver writes to
3888 * each DVA, including the one(s) that are not on vdevs with dirty
3891 * 2. With nested replication, which happens when we have a
3892 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3893 * For example, given mirror(replacing(A+B), C), it's likely that
3894 * only A is out of date (it's the new device). In this case, we'll
3895 * read from C, then use the data to resilver A+B -- but we don't
3896 * actually want to resilver B, just A. The top-level mirror has no
3897 * way to know this, so instead we just discard unnecessary repairs
3898 * as we work our way down the vdev tree.
3900 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3901 * The same logic applies to any form of nested replication: ditto
3902 * + mirror, RAID-Z + replacing, etc.
3904 * However, indirect vdevs point off to other vdevs which may have
3905 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3906 * will be properly bypassed instead.
3908 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3909 * a dRAID spare vdev. For example, when a dRAID spare is first
3910 * used, its spare blocks need to be written to but the leaf vdev's
3911 * of such blocks can have empty DTL_PARTIAL.
3913 * There seemed no clean way to allow such writes while bypassing
3914 * spurious ones. At this point, just avoid all bypassing for dRAID
3917 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3918 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3919 zio
->io_txg
!= 0 && /* not a delegated i/o */
3920 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3921 vd
->vdev_top
->vdev_ops
!= &vdev_draid_ops
&&
3922 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3923 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3924 zio_vdev_io_bypass(zio
);
3929 * Select the next best leaf I/O to process. Distributed spares are
3930 * excluded since they dispatch the I/O directly to a leaf vdev after
3931 * applying the dRAID mapping.
3933 if (vd
->vdev_ops
->vdev_op_leaf
&&
3934 vd
->vdev_ops
!= &vdev_draid_spare_ops
&&
3935 (zio
->io_type
== ZIO_TYPE_READ
||
3936 zio
->io_type
== ZIO_TYPE_WRITE
||
3937 zio
->io_type
== ZIO_TYPE_TRIM
)) {
3939 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3942 if (!vdev_accessible(vd
, zio
)) {
3943 zio
->io_error
= SET_ERROR(ENXIO
);
3947 zio
->io_delay
= gethrtime();
3950 vd
->vdev_ops
->vdev_op_io_start(zio
);
3955 zio_vdev_io_done(zio_t
*zio
)
3957 vdev_t
*vd
= zio
->io_vd
;
3958 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3959 boolean_t unexpected_error
= B_FALSE
;
3961 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3965 ASSERT(zio
->io_type
== ZIO_TYPE_READ
||
3966 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
);
3969 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3971 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3972 vd
->vdev_ops
!= &vdev_draid_spare_ops
) {
3973 vdev_queue_io_done(zio
);
3975 if (zio_injection_enabled
&& zio
->io_error
== 0)
3976 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3979 if (zio_injection_enabled
&& zio
->io_error
== 0)
3980 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3982 if (zio
->io_error
&& zio
->io_type
!= ZIO_TYPE_TRIM
) {
3983 if (!vdev_accessible(vd
, zio
)) {
3984 zio
->io_error
= SET_ERROR(ENXIO
);
3986 unexpected_error
= B_TRUE
;
3991 ops
->vdev_op_io_done(zio
);
3993 if (unexpected_error
&& vd
->vdev_remove_wanted
== B_FALSE
)
3994 VERIFY(vdev_probe(vd
, zio
) == NULL
);
4000 * This function is used to change the priority of an existing zio that is
4001 * currently in-flight. This is used by the arc to upgrade priority in the
4002 * event that a demand read is made for a block that is currently queued
4003 * as a scrub or async read IO. Otherwise, the high priority read request
4004 * would end up having to wait for the lower priority IO.
4007 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
4009 zio_t
*cio
, *cio_next
;
4010 zio_link_t
*zl
= NULL
;
4012 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
4014 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
4015 vdev_queue_change_io_priority(pio
, priority
);
4017 pio
->io_priority
= priority
;
4020 mutex_enter(&pio
->io_lock
);
4021 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
4022 cio_next
= zio_walk_children(pio
, &zl
);
4023 zio_change_priority(cio
, priority
);
4025 mutex_exit(&pio
->io_lock
);
4029 * For non-raidz ZIOs, we can just copy aside the bad data read from the
4030 * disk, and use that to finish the checksum ereport later.
4033 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
4034 const abd_t
*good_buf
)
4036 /* no processing needed */
4037 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
4041 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
)
4043 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
4045 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
4047 zcr
->zcr_cbinfo
= zio
->io_size
;
4048 zcr
->zcr_cbdata
= abd
;
4049 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
4050 zcr
->zcr_free
= zio_abd_free
;
4054 zio_vdev_io_assess(zio_t
*zio
)
4056 vdev_t
*vd
= zio
->io_vd
;
4058 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
4062 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
4063 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
4065 if (zio
->io_vsd
!= NULL
) {
4066 zio
->io_vsd_ops
->vsd_free(zio
);
4070 if (zio_injection_enabled
&& zio
->io_error
== 0)
4071 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
4074 * If the I/O failed, determine whether we should attempt to retry it.
4076 * On retry, we cut in line in the issue queue, since we don't want
4077 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4079 if (zio
->io_error
&& vd
== NULL
&&
4080 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
4081 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
4082 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
4084 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
| ZIO_FLAG_DONT_AGGREGATE
;
4085 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
4086 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
4087 zio_requeue_io_start_cut_in_line
);
4092 * If we got an error on a leaf device, convert it to ENXIO
4093 * if the device is not accessible at all.
4095 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
4096 !vdev_accessible(vd
, zio
))
4097 zio
->io_error
= SET_ERROR(ENXIO
);
4100 * If we can't write to an interior vdev (mirror or RAID-Z),
4101 * set vdev_cant_write so that we stop trying to allocate from it.
4103 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
4104 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
4105 vdev_dbgmsg(vd
, "zio_vdev_io_assess(zio=%px) setting "
4106 "cant_write=TRUE due to write failure with ENXIO",
4108 vd
->vdev_cant_write
= B_TRUE
;
4112 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4113 * attempts will ever succeed. In this case we set a persistent
4114 * boolean flag so that we don't bother with it in the future.
4116 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
4117 zio
->io_type
== ZIO_TYPE_IOCTL
&&
4118 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
4119 vd
->vdev_nowritecache
= B_TRUE
;
4122 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4128 zio_vdev_io_reissue(zio_t
*zio
)
4130 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
4131 ASSERT(zio
->io_error
== 0);
4133 zio
->io_stage
>>= 1;
4137 zio_vdev_io_redone(zio_t
*zio
)
4139 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
4141 zio
->io_stage
>>= 1;
4145 zio_vdev_io_bypass(zio_t
*zio
)
4147 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
4148 ASSERT(zio
->io_error
== 0);
4150 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
4151 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
4155 * ==========================================================================
4156 * Encrypt and store encryption parameters
4157 * ==========================================================================
4162 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4163 * managing the storage of encryption parameters and passing them to the
4164 * lower-level encryption functions.
4167 zio_encrypt(zio_t
*zio
)
4169 zio_prop_t
*zp
= &zio
->io_prop
;
4170 spa_t
*spa
= zio
->io_spa
;
4171 blkptr_t
*bp
= zio
->io_bp
;
4172 uint64_t psize
= BP_GET_PSIZE(bp
);
4173 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
4174 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
4175 void *enc_buf
= NULL
;
4177 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4178 uint8_t iv
[ZIO_DATA_IV_LEN
];
4179 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4180 boolean_t no_crypt
= B_FALSE
;
4182 /* the root zio already encrypted the data */
4183 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
4186 /* only ZIL blocks are re-encrypted on rewrite */
4187 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
4190 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
4191 BP_SET_CRYPT(bp
, B_FALSE
);
4195 /* if we are doing raw encryption set the provided encryption params */
4196 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
4197 ASSERT0(BP_GET_LEVEL(bp
));
4198 BP_SET_CRYPT(bp
, B_TRUE
);
4199 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
4200 if (ot
!= DMU_OT_OBJSET
)
4201 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
4203 /* dnode blocks must be written out in the provided byteorder */
4204 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
4205 ot
== DMU_OT_DNODE
) {
4206 void *bswap_buf
= zio_buf_alloc(psize
);
4207 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
4209 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4210 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
4211 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
4214 abd_take_ownership_of_buf(babd
, B_TRUE
);
4215 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
4218 if (DMU_OT_IS_ENCRYPTED(ot
))
4219 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
4223 /* indirect blocks only maintain a cksum of the lower level MACs */
4224 if (BP_GET_LEVEL(bp
) > 0) {
4225 BP_SET_CRYPT(bp
, B_TRUE
);
4226 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
4227 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
4229 zio_crypt_encode_mac_bp(bp
, mac
);
4234 * Objset blocks are a special case since they have 2 256-bit MACs
4235 * embedded within them.
4237 if (ot
== DMU_OT_OBJSET
) {
4238 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
4239 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4240 BP_SET_CRYPT(bp
, B_TRUE
);
4241 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
4242 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
4246 /* unencrypted object types are only authenticated with a MAC */
4247 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
4248 BP_SET_CRYPT(bp
, B_TRUE
);
4249 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
4250 zio
->io_abd
, psize
, mac
));
4251 zio_crypt_encode_mac_bp(bp
, mac
);
4256 * Later passes of sync-to-convergence may decide to rewrite data
4257 * in place to avoid more disk reallocations. This presents a problem
4258 * for encryption because this constitutes rewriting the new data with
4259 * the same encryption key and IV. However, this only applies to blocks
4260 * in the MOS (particularly the spacemaps) and we do not encrypt the
4261 * MOS. We assert that the zio is allocating or an intent log write
4264 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
4265 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
4266 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
4267 ASSERT3U(psize
, !=, 0);
4269 enc_buf
= zio_buf_alloc(psize
);
4270 eabd
= abd_get_from_buf(enc_buf
, psize
);
4271 abd_take_ownership_of_buf(eabd
, B_TRUE
);
4274 * For an explanation of what encryption parameters are stored
4275 * where, see the block comment in zio_crypt.c.
4277 if (ot
== DMU_OT_INTENT_LOG
) {
4278 zio_crypt_decode_params_bp(bp
, salt
, iv
);
4280 BP_SET_CRYPT(bp
, B_TRUE
);
4283 /* Perform the encryption. This should not fail */
4284 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
4285 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
4286 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
4288 /* encode encryption metadata into the bp */
4289 if (ot
== DMU_OT_INTENT_LOG
) {
4291 * ZIL blocks store the MAC in the embedded checksum, so the
4292 * transform must always be applied.
4294 zio_crypt_encode_mac_zil(enc_buf
, mac
);
4295 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4297 BP_SET_CRYPT(bp
, B_TRUE
);
4298 zio_crypt_encode_params_bp(bp
, salt
, iv
);
4299 zio_crypt_encode_mac_bp(bp
, mac
);
4302 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
4305 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4313 * ==========================================================================
4314 * Generate and verify checksums
4315 * ==========================================================================
4318 zio_checksum_generate(zio_t
*zio
)
4320 blkptr_t
*bp
= zio
->io_bp
;
4321 enum zio_checksum checksum
;
4325 * This is zio_write_phys().
4326 * We're either generating a label checksum, or none at all.
4328 checksum
= zio
->io_prop
.zp_checksum
;
4330 if (checksum
== ZIO_CHECKSUM_OFF
)
4333 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4335 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4336 ASSERT(!IO_IS_ALLOCATING(zio
));
4337 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4339 checksum
= BP_GET_CHECKSUM(bp
);
4343 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4349 zio_checksum_verify(zio_t
*zio
)
4351 zio_bad_cksum_t info
;
4352 blkptr_t
*bp
= zio
->io_bp
;
4355 ASSERT(zio
->io_vd
!= NULL
);
4359 * This is zio_read_phys().
4360 * We're either verifying a label checksum, or nothing at all.
4362 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4365 ASSERT3U(zio
->io_prop
.zp_checksum
, ==, ZIO_CHECKSUM_LABEL
);
4368 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4369 zio
->io_error
= error
;
4370 if (error
== ECKSUM
&&
4371 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4372 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4373 zio
->io_vd
->vdev_stat
.vs_checksum_errors
++;
4374 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4375 (void) zfs_ereport_start_checksum(zio
->io_spa
,
4376 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4377 zio
->io_offset
, zio
->io_size
, &info
);
4385 * Called by RAID-Z to ensure we don't compute the checksum twice.
4388 zio_checksum_verified(zio_t
*zio
)
4390 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4394 * ==========================================================================
4395 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4396 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4397 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4398 * indicate errors that are specific to one I/O, and most likely permanent.
4399 * Any other error is presumed to be worse because we weren't expecting it.
4400 * ==========================================================================
4403 zio_worst_error(int e1
, int e2
)
4405 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4408 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4409 if (e1
== zio_error_rank
[r1
])
4412 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4413 if (e2
== zio_error_rank
[r2
])
4416 return (r1
> r2
? e1
: e2
);
4420 * ==========================================================================
4422 * ==========================================================================
4425 zio_ready(zio_t
*zio
)
4427 blkptr_t
*bp
= zio
->io_bp
;
4428 zio_t
*pio
, *pio_next
;
4429 zio_link_t
*zl
= NULL
;
4431 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4436 if (zio
->io_ready
) {
4437 ASSERT(IO_IS_ALLOCATING(zio
));
4438 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4439 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4440 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4445 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4446 zio
->io_bp_copy
= *bp
;
4448 if (zio
->io_error
!= 0) {
4449 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4451 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4452 ASSERT(IO_IS_ALLOCATING(zio
));
4453 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4454 ASSERT(zio
->io_metaslab_class
!= NULL
);
4457 * We were unable to allocate anything, unreserve and
4458 * issue the next I/O to allocate.
4460 metaslab_class_throttle_unreserve(
4461 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4462 zio
->io_allocator
, zio
);
4463 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4467 mutex_enter(&zio
->io_lock
);
4468 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4469 pio
= zio_walk_parents(zio
, &zl
);
4470 mutex_exit(&zio
->io_lock
);
4473 * As we notify zio's parents, new parents could be added.
4474 * New parents go to the head of zio's io_parent_list, however,
4475 * so we will (correctly) not notify them. The remainder of zio's
4476 * io_parent_list, from 'pio_next' onward, cannot change because
4477 * all parents must wait for us to be done before they can be done.
4479 for (; pio
!= NULL
; pio
= pio_next
) {
4480 pio_next
= zio_walk_parents(zio
, &zl
);
4481 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4484 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4485 if (bp
!= NULL
&& BP_IS_GANG(bp
)) {
4486 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4488 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4489 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4493 if (zio_injection_enabled
&&
4494 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4495 zio_handle_ignored_writes(zio
);
4501 * Update the allocation throttle accounting.
4504 zio_dva_throttle_done(zio_t
*zio
)
4506 zio_t
*lio __maybe_unused
= zio
->io_logical
;
4507 zio_t
*pio
= zio_unique_parent(zio
);
4508 vdev_t
*vd
= zio
->io_vd
;
4509 int flags
= METASLAB_ASYNC_ALLOC
;
4511 ASSERT3P(zio
->io_bp
, !=, NULL
);
4512 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4513 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4514 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4516 ASSERT3P(vd
, ==, vd
->vdev_top
);
4517 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4518 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4519 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4520 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4521 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4524 * Parents of gang children can have two flavors -- ones that
4525 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4526 * and ones that allocated the constituent blocks. The allocation
4527 * throttle needs to know the allocating parent zio so we must find
4530 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4532 * If our parent is a rewrite gang child then our grandparent
4533 * would have been the one that performed the allocation.
4535 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4536 pio
= zio_unique_parent(pio
);
4537 flags
|= METASLAB_GANG_CHILD
;
4540 ASSERT(IO_IS_ALLOCATING(pio
));
4541 ASSERT3P(zio
, !=, zio
->io_logical
);
4542 ASSERT(zio
->io_logical
!= NULL
);
4543 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4544 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4545 ASSERT(zio
->io_metaslab_class
!= NULL
);
4547 mutex_enter(&pio
->io_lock
);
4548 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4549 pio
->io_allocator
, B_TRUE
);
4550 mutex_exit(&pio
->io_lock
);
4552 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4553 pio
->io_allocator
, pio
);
4556 * Call into the pipeline to see if there is more work that
4557 * needs to be done. If there is work to be done it will be
4558 * dispatched to another taskq thread.
4560 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4564 zio_done(zio_t
*zio
)
4567 * Always attempt to keep stack usage minimal here since
4568 * we can be called recursively up to 19 levels deep.
4570 const uint64_t psize
= zio
->io_size
;
4571 zio_t
*pio
, *pio_next
;
4572 zio_link_t
*zl
= NULL
;
4575 * If our children haven't all completed,
4576 * wait for them and then repeat this pipeline stage.
4578 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4583 * If the allocation throttle is enabled, then update the accounting.
4584 * We only track child I/Os that are part of an allocating async
4585 * write. We must do this since the allocation is performed
4586 * by the logical I/O but the actual write is done by child I/Os.
4588 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4589 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4590 ASSERT(zio
->io_metaslab_class
!= NULL
);
4591 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4592 zio_dva_throttle_done(zio
);
4596 * If the allocation throttle is enabled, verify that
4597 * we have decremented the refcounts for every I/O that was throttled.
4599 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4600 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4601 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4602 ASSERT(zio
->io_bp
!= NULL
);
4604 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4606 VERIFY(zfs_refcount_not_held(&zio
->io_metaslab_class
->
4607 mc_allocator
[zio
->io_allocator
].mca_alloc_slots
, zio
));
4611 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4612 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4613 ASSERT(zio
->io_children
[c
][w
] == 0);
4615 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4616 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4617 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4618 ASSERT(memcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4619 sizeof (blkptr_t
)) == 0 ||
4620 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4621 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4622 zio
->io_bp_override
== NULL
&&
4623 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4624 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4625 BP_GET_NDVAS(zio
->io_bp
));
4626 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4627 (BP_COUNT_GANG(zio
->io_bp
) ==
4628 BP_GET_NDVAS(zio
->io_bp
)));
4630 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4631 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4635 * If there were child vdev/gang/ddt errors, they apply to us now.
4637 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4638 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4639 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4642 * If the I/O on the transformed data was successful, generate any
4643 * checksum reports now while we still have the transformed data.
4645 if (zio
->io_error
== 0) {
4646 while (zio
->io_cksum_report
!= NULL
) {
4647 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4648 uint64_t align
= zcr
->zcr_align
;
4649 uint64_t asize
= P2ROUNDUP(psize
, align
);
4650 abd_t
*adata
= zio
->io_abd
;
4652 if (adata
!= NULL
&& asize
!= psize
) {
4653 adata
= abd_alloc(asize
, B_TRUE
);
4654 abd_copy(adata
, zio
->io_abd
, psize
);
4655 abd_zero_off(adata
, psize
, asize
- psize
);
4658 zio
->io_cksum_report
= zcr
->zcr_next
;
4659 zcr
->zcr_next
= NULL
;
4660 zcr
->zcr_finish(zcr
, adata
);
4661 zfs_ereport_free_checksum(zcr
);
4663 if (adata
!= NULL
&& asize
!= psize
)
4668 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4670 vdev_stat_update(zio
, psize
);
4673 * If this I/O is attached to a particular vdev is slow, exceeding
4674 * 30 seconds to complete, post an error described the I/O delay.
4675 * We ignore these errors if the device is currently unavailable.
4677 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4678 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4680 * We want to only increment our slow IO counters if
4681 * the IO is valid (i.e. not if the drive is removed).
4683 * zfs_ereport_post() will also do these checks, but
4684 * it can also ratelimit and have other failures, so we
4685 * need to increment the slow_io counters independent
4688 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4689 zio
->io_spa
, zio
->io_vd
, zio
)) {
4690 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4691 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4692 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4694 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4695 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4701 if (zio
->io_error
) {
4703 * If this I/O is attached to a particular vdev,
4704 * generate an error message describing the I/O failure
4705 * at the block level. We ignore these errors if the
4706 * device is currently unavailable.
4708 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4709 !vdev_is_dead(zio
->io_vd
)) {
4710 int ret
= zfs_ereport_post(FM_EREPORT_ZFS_IO
,
4711 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
, zio
, 0);
4712 if (ret
!= EALREADY
) {
4713 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4714 if (zio
->io_type
== ZIO_TYPE_READ
)
4715 zio
->io_vd
->vdev_stat
.vs_read_errors
++;
4716 else if (zio
->io_type
== ZIO_TYPE_WRITE
)
4717 zio
->io_vd
->vdev_stat
.vs_write_errors
++;
4718 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4722 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4723 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4724 zio
== zio
->io_logical
) {
4726 * For logical I/O requests, tell the SPA to log the
4727 * error and generate a logical data ereport.
4729 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
,
4730 &zio
->io_bp
->blk_birth
);
4731 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA
,
4732 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
, 0);
4736 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4738 * Determine whether zio should be reexecuted. This will
4739 * propagate all the way to the root via zio_notify_parent().
4741 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4742 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4744 if (IO_IS_ALLOCATING(zio
) &&
4745 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4746 if (zio
->io_error
!= ENOSPC
)
4747 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4749 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4752 if ((zio
->io_type
== ZIO_TYPE_READ
||
4753 zio
->io_type
== ZIO_TYPE_FREE
) &&
4754 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4755 zio
->io_error
== ENXIO
&&
4756 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4757 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4758 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4760 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4761 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4764 * Here is a possibly good place to attempt to do
4765 * either combinatorial reconstruction or error correction
4766 * based on checksums. It also might be a good place
4767 * to send out preliminary ereports before we suspend
4773 * If there were logical child errors, they apply to us now.
4774 * We defer this until now to avoid conflating logical child
4775 * errors with errors that happened to the zio itself when
4776 * updating vdev stats and reporting FMA events above.
4778 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4780 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4781 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4782 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4783 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4785 zio_gang_tree_free(&zio
->io_gang_tree
);
4788 * Godfather I/Os should never suspend.
4790 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4791 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4792 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4794 if (zio
->io_reexecute
) {
4796 * This is a logical I/O that wants to reexecute.
4798 * Reexecute is top-down. When an i/o fails, if it's not
4799 * the root, it simply notifies its parent and sticks around.
4800 * The parent, seeing that it still has children in zio_done(),
4801 * does the same. This percolates all the way up to the root.
4802 * The root i/o will reexecute or suspend the entire tree.
4804 * This approach ensures that zio_reexecute() honors
4805 * all the original i/o dependency relationships, e.g.
4806 * parents not executing until children are ready.
4808 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4810 zio
->io_gang_leader
= NULL
;
4812 mutex_enter(&zio
->io_lock
);
4813 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4814 mutex_exit(&zio
->io_lock
);
4817 * "The Godfather" I/O monitors its children but is
4818 * not a true parent to them. It will track them through
4819 * the pipeline but severs its ties whenever they get into
4820 * trouble (e.g. suspended). This allows "The Godfather"
4821 * I/O to return status without blocking.
4824 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4826 zio_link_t
*remove_zl
= zl
;
4827 pio_next
= zio_walk_parents(zio
, &zl
);
4829 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4830 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4831 zio_remove_child(pio
, zio
, remove_zl
);
4833 * This is a rare code path, so we don't
4834 * bother with "next_to_execute".
4836 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4841 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4843 * We're not a root i/o, so there's nothing to do
4844 * but notify our parent. Don't propagate errors
4845 * upward since we haven't permanently failed yet.
4847 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4848 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4850 * This is a rare code path, so we don't bother with
4851 * "next_to_execute".
4853 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4854 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4856 * We'd fail again if we reexecuted now, so suspend
4857 * until conditions improve (e.g. device comes online).
4859 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4862 * Reexecution is potentially a huge amount of work.
4863 * Hand it off to the otherwise-unused claim taskq.
4865 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4866 spa_taskq_dispatch_ent(zio
->io_spa
,
4867 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4868 zio_reexecute
, zio
, 0, &zio
->io_tqent
);
4873 ASSERT(list_is_empty(&zio
->io_child_list
));
4874 ASSERT(zio
->io_reexecute
== 0);
4875 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4878 * Report any checksum errors, since the I/O is complete.
4880 while (zio
->io_cksum_report
!= NULL
) {
4881 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4882 zio
->io_cksum_report
= zcr
->zcr_next
;
4883 zcr
->zcr_next
= NULL
;
4884 zcr
->zcr_finish(zcr
, NULL
);
4885 zfs_ereport_free_checksum(zcr
);
4888 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4889 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4890 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4891 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4895 * It is the responsibility of the done callback to ensure that this
4896 * particular zio is no longer discoverable for adoption, and as
4897 * such, cannot acquire any new parents.
4902 mutex_enter(&zio
->io_lock
);
4903 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4904 mutex_exit(&zio
->io_lock
);
4907 * We are done executing this zio. We may want to execute a parent
4908 * next. See the comment in zio_notify_parent().
4910 zio_t
*next_to_execute
= NULL
;
4912 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4913 zio_link_t
*remove_zl
= zl
;
4914 pio_next
= zio_walk_parents(zio
, &zl
);
4915 zio_remove_child(pio
, zio
, remove_zl
);
4916 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
4919 if (zio
->io_waiter
!= NULL
) {
4920 mutex_enter(&zio
->io_lock
);
4921 zio
->io_executor
= NULL
;
4922 cv_broadcast(&zio
->io_cv
);
4923 mutex_exit(&zio
->io_lock
);
4928 return (next_to_execute
);
4932 * ==========================================================================
4933 * I/O pipeline definition
4934 * ==========================================================================
4936 static zio_pipe_stage_t
*zio_pipeline
[] = {
4944 zio_checksum_generate
,
4961 zio_checksum_verify
,
4969 * Compare two zbookmark_phys_t's to see which we would reach first in a
4970 * pre-order traversal of the object tree.
4972 * This is simple in every case aside from the meta-dnode object. For all other
4973 * objects, we traverse them in order (object 1 before object 2, and so on).
4974 * However, all of these objects are traversed while traversing object 0, since
4975 * the data it points to is the list of objects. Thus, we need to convert to a
4976 * canonical representation so we can compare meta-dnode bookmarks to
4977 * non-meta-dnode bookmarks.
4979 * We do this by calculating "equivalents" for each field of the zbookmark.
4980 * zbookmarks outside of the meta-dnode use their own object and level, and
4981 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4982 * blocks this bookmark refers to) by multiplying their blkid by their span
4983 * (the number of L0 blocks contained within one block at their level).
4984 * zbookmarks inside the meta-dnode calculate their object equivalent
4985 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4986 * level + 1<<31 (any value larger than a level could ever be) for their level.
4987 * This causes them to always compare before a bookmark in their object
4988 * equivalent, compare appropriately to bookmarks in other objects, and to
4989 * compare appropriately to other bookmarks in the meta-dnode.
4992 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4993 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4996 * These variables represent the "equivalent" values for the zbookmark,
4997 * after converting zbookmarks inside the meta dnode to their
4998 * normal-object equivalents.
5000 uint64_t zb1obj
, zb2obj
;
5001 uint64_t zb1L0
, zb2L0
;
5002 uint64_t zb1level
, zb2level
;
5004 if (zb1
->zb_object
== zb2
->zb_object
&&
5005 zb1
->zb_level
== zb2
->zb_level
&&
5006 zb1
->zb_blkid
== zb2
->zb_blkid
)
5009 IMPLY(zb1
->zb_level
> 0, ibs1
>= SPA_MINBLOCKSHIFT
);
5010 IMPLY(zb2
->zb_level
> 0, ibs2
>= SPA_MINBLOCKSHIFT
);
5013 * BP_SPANB calculates the span in blocks.
5015 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
5016 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
5018 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
5019 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
5021 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
5023 zb1obj
= zb1
->zb_object
;
5024 zb1level
= zb1
->zb_level
;
5027 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
5028 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
5030 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
5032 zb2obj
= zb2
->zb_object
;
5033 zb2level
= zb2
->zb_level
;
5036 /* Now that we have a canonical representation, do the comparison. */
5037 if (zb1obj
!= zb2obj
)
5038 return (zb1obj
< zb2obj
? -1 : 1);
5039 else if (zb1L0
!= zb2L0
)
5040 return (zb1L0
< zb2L0
? -1 : 1);
5041 else if (zb1level
!= zb2level
)
5042 return (zb1level
> zb2level
? -1 : 1);
5044 * This can (theoretically) happen if the bookmarks have the same object
5045 * and level, but different blkids, if the block sizes are not the same.
5046 * There is presently no way to change the indirect block sizes
5052 * This function checks the following: given that last_block is the place that
5053 * our traversal stopped last time, does that guarantee that we've visited
5054 * every node under subtree_root? Therefore, we can't just use the raw output
5055 * of zbookmark_compare. We have to pass in a modified version of
5056 * subtree_root; by incrementing the block id, and then checking whether
5057 * last_block is before or equal to that, we can tell whether or not having
5058 * visited last_block implies that all of subtree_root's children have been
5062 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
5063 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
5065 zbookmark_phys_t mod_zb
= *subtree_root
;
5067 ASSERT0(last_block
->zb_level
);
5069 /* The objset_phys_t isn't before anything. */
5074 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5075 * data block size in sectors, because that variable is only used if
5076 * the bookmark refers to a block in the meta-dnode. Since we don't
5077 * know without examining it what object it refers to, and there's no
5078 * harm in passing in this value in other cases, we always pass it in.
5080 * We pass in 0 for the indirect block size shift because zb2 must be
5081 * level 0. The indirect block size is only used to calculate the span
5082 * of the bookmark, but since the bookmark must be level 0, the span is
5083 * always 1, so the math works out.
5085 * If you make changes to how the zbookmark_compare code works, be sure
5086 * to make sure that this code still works afterwards.
5088 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
5089 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
5094 * This function is similar to zbookmark_subtree_completed(), but returns true
5095 * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5098 zbookmark_subtree_tbd(const dnode_phys_t
*dnp
,
5099 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
5101 ASSERT0(last_block
->zb_level
);
5104 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
5105 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, subtree_root
,
5109 EXPORT_SYMBOL(zio_type_name
);
5110 EXPORT_SYMBOL(zio_buf_alloc
);
5111 EXPORT_SYMBOL(zio_data_buf_alloc
);
5112 EXPORT_SYMBOL(zio_buf_free
);
5113 EXPORT_SYMBOL(zio_data_buf_free
);
5115 ZFS_MODULE_PARAM(zfs_zio
, zio_
, slow_io_ms
, INT
, ZMOD_RW
,
5116 "Max I/O completion time (milliseconds) before marking it as slow");
5118 ZFS_MODULE_PARAM(zfs_zio
, zio_
, requeue_io_start_cut_in_line
, INT
, ZMOD_RW
,
5119 "Prioritize requeued I/O");
5121 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_deferred_free
, UINT
, ZMOD_RW
,
5122 "Defer frees starting in this pass");
5124 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_dont_compress
, UINT
, ZMOD_RW
,
5125 "Don't compress starting in this pass");
5127 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_rewrite
, UINT
, ZMOD_RW
,
5128 "Rewrite new bps starting in this pass");
5130 ZFS_MODULE_PARAM(zfs_zio
, zio_
, dva_throttle_enabled
, INT
, ZMOD_RW
,
5131 "Throttle block allocations in the ZIO pipeline");
5133 ZFS_MODULE_PARAM(zfs_zio
, zio_
, deadman_log_all
, INT
, ZMOD_RW
,
5134 "Log all slow ZIOs, not just those with vdevs");